diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index b653b1f823..1e24293645 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -467,6 +467,7 @@ jobs: run: | cmake -B build -S . \ -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \ + -DGGML_HIP_ROCWMMA_FATTN=ON \ -DGGML_HIP=ON cmake --build build --config Release -j $(nproc) @@ -476,6 +477,7 @@ jobs: cmake -B build2 -S . \ -DCMAKE_C_COMPILER=hipcc \ -DCMAKE_CXX_COMPILER=hipcc \ + -DGGML_HIP_ROCWMMA_FATTN=ON \ -DGGML_HIP=ON cmake --build build2 --config Release -j $(nproc) @@ -772,7 +774,7 @@ jobs: env: OPENBLAS_VERSION: 0.3.23 SDE_VERSION: 9.33.0-2024-01-07 - VULKAN_VERSION: 1.3.261.1 + VULKAN_VERSION: 1.4.304.1 strategy: matrix: @@ -1202,6 +1204,11 @@ jobs: id: checkout uses: actions/checkout@v4 + - name: Clone rocWMMA repository + id: clone_rocwmma + run: | + git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1 + - name: Install id: depends run: | @@ -1231,8 +1238,10 @@ jobs: cmake -G "Unix Makefiles" -B build -S . ` -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" ` -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" ` + -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/" ` -DCMAKE_BUILD_TYPE=Release ` -DGGML_HIP=ON ` + -DGGML_HIP_ROCWMMA_FATTN=ON ` -DGGML_RPC=ON cmake --build build -j ${env:NUMBER_OF_PROCESSORS} @@ -1251,6 +1260,11 @@ jobs: with: fetch-depth: 0 + - name: Clone rocWMMA repository + id: clone_rocwmma + run: | + git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1 + - name: ccache uses: hendrikmuhs/ccache-action@v1.2.16 with: @@ -1280,8 +1294,10 @@ jobs: cmake -G "Unix Makefiles" -B build -S . ` -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" ` -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" ` + -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/" ` -DCMAKE_BUILD_TYPE=Release ` -DAMDGPU_TARGETS=${{ matrix.gpu_target }} ` + -DGGML_HIP_ROCWMMA_FATTN=ON ` -DGGML_HIP=ON ` -DGGML_RPC=ON cmake --build build -j ${env:NUMBER_OF_PROCESSORS} diff --git a/AUTHORS b/AUTHORS index 6796b29413..0af9f44ad4 100644 --- a/AUTHORS +++ b/AUTHORS @@ -1,4 +1,4 @@ -# date: Tue Feb 4 13:04:05 EET 2025 +# date: Sat Mar 8 18:23:52 EET 2025 # this file is auto-generated by scripts/gen-authors.sh 0cc4m @@ -8,10 +8,12 @@ 3ooabkhxtn <31479382+3ooabkhxtn@users.noreply.github.com> 44670 <44670@users.noreply.github.com> 65a <10104049+65a@users.noreply.github.com> +708-145 <40387547+708-145@users.noreply.github.com> AN Long AT Aarni Koskela Aaron Miller +Aaron Teo <57927438+taronaeo@users.noreply.github.com> Aaryaman Vasishta Abheek Gulati Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com> @@ -20,6 +22,7 @@ Adithya Balaji AdithyanI Adrian Adrian Hesketh +Adrian Kretz Adrien Gallouët Adrien Gallouët Ahmad Tameem <113388789+Tameem-10xE@users.noreply.github.com> @@ -28,15 +31,18 @@ AidanBeltonS <87009434+AidanBeltonS@users.noreply.github.com> AidanBeltonS Aisuko Akarshan Biswas +Akarshan Biswas Akarshan Biswas Al Mochkin <14274697+amochkin@users.noreply.github.com> Albert Jin Alberto <57916483+albbus-stack@users.noreply.github.com> Alberto Cabrera Pérez Alberto Cabrera Pérez +Aleksei Nikiforov <103434461+AlekseiNikiforovIBM@users.noreply.github.com> Alex Alex Azarov Alex Azarov +Alex Brooks Alex Klinkhamer Alex Klinkhamer Alex Nguyen @@ -67,6 +73,7 @@ Andrew Minh Nguyen <40281306+amqdn@users.noreply.github.com> Andy Salerno Andy Tai Anthony Van de Gejuchte +Antoine Viallon Antonis Makropoulos Arik Poznanski Armen Kaleshian @@ -83,6 +90,7 @@ Atsushi Tatsuma Austin <77757836+teleprint-me@users.noreply.github.com> AustinMroz BADR +BB-fat <45072480+BB-fat@users.noreply.github.com> Bach Le Bailey Chittle <39804642+bachittle@users.noreply.github.com> BarfingLemurs <128182951+BarfingLemurs@users.noreply.github.com> @@ -101,6 +109,7 @@ Bert Wagner Billel Mokeddem Bingan <70050083+binganao@users.noreply.github.com> Bjarke Viksøe <164612031+bviksoe@users.noreply.github.com> +Bodhi <3882561+BodhiHu@users.noreply.github.com> Bodo Graumann Bono Lv Borislav Stanimirov @@ -128,6 +137,7 @@ CentricStorm Chad Brewbaker Changyeon Kim Chao Jiang +Charles Duffy Charles Xu <63788048+chaxu01@users.noreply.github.com> Charles Xu Chen Xi @@ -139,12 +149,14 @@ Chris Kuehl Christian Demsar Christian Demsar Christian Falch <875252+chrfalch@users.noreply.github.com> +Christian Fillion Christian Kastner Christian Kögler Christian Köhnenkamp Christian Zhou-Zheng <59622928+christianazinn@users.noreply.github.com> Christopher Nielsen <62156882+mascguy@users.noreply.github.com> Clark Saben <76020733+csaben@users.noreply.github.com> +Clauszy Clint Herron Conrad Kramer Corentin REGAL @@ -163,6 +175,7 @@ Daniel Hiltgen Daniel Illescas Romero Daniel Kleine <53251018+d-kleine@users.noreply.github.com> Daniele <57776841+daniandtheweb@users.noreply.github.com> +Danny Milosavljevic DannyDaemonic Dat Quoc Nguyen <2412555+datquocnguyen@users.noreply.github.com> Dave @@ -170,6 +183,7 @@ Dave Airlie Dave Airlie Dave Della Costa David Friehs +David Huang <1969802+hjc4869@users.noreply.github.com> David Kennedy David Pflug David Renshaw @@ -236,6 +250,7 @@ Felix Finn Voorhees Firat FirstTimeEZ <179362031+FirstTimeEZ@users.noreply.github.com> +Florent BENOIT Folko-Ven <71110216+Folko-Ven@users.noreply.github.com> Foul-Tarnished <107711110+Foul-Tarnished@users.noreply.github.com> Francisco Melo <43780565+francis2tm@users.noreply.github.com> @@ -254,6 +269,7 @@ Gary Mulder Gavin Zhao Genkagaku.GPT Georgi Gerganov +Gian-Carlo Pascutto Gilad S Gilad S. <7817232+giladgd@users.noreply.github.com> Giuseppe Scrivano @@ -267,7 +283,9 @@ Guspan Tanadi <36249910+guspan-tanadi@users.noreply.github.com> Gustavo Rocha Dias <91472747+gustrd@users.noreply.github.com> Haggai Nuchi Halalaluyafail3 <55773281+Halalaluyafail3@users.noreply.github.com> +Hale Chan Hamdoud Hakem <90524568+hamdoudhakem@users.noreply.github.com> +Han Yin HanishKVC Haohui Mai Haoxiang Fei @@ -278,6 +296,7 @@ Haus1 Henk Poley Henri Vasserman Henrik Forstén +Henry Linjamäki Herman Semenov Hesen Peng HimariO @@ -307,6 +326,7 @@ Ivan Ivan Filipov <159561759+vanaka11@users.noreply.github.com> Ivan Komarov Ivan Stepanov +JC <43374599+MrSMlT@users.noreply.github.com> JFLFY2255 JH23X <165871467+JH23X@users.noreply.github.com> Jack Mousseau @@ -325,6 +345,7 @@ Jan Ploski Jannis Schönleber Jared Van Bortel Jared Van Bortel +Jason C.H Jason McCartney Jason Stillerman Jean-Christophe Hoelt @@ -342,6 +363,7 @@ Jiahao Li Jian Liao JidongZhang-THU <1119708529@qq.com> Jinwoo Jeong <33892306+williamjeong2@users.noreply.github.com> +Jinyang He Jiří Podivín <66251151+jpodivin@users.noreply.github.com> Jiří Sejkora Joan Fontanals @@ -379,6 +401,7 @@ Justine Tunney Juuso Alasuutari KASR Kamil Tomšík +Kante Yin Karol Kontny <82021046+kkontny@users.noreply.github.com> Karsten Weiss Karthick @@ -419,6 +442,7 @@ LoganDark Loïc Carrère LostRuins <39025047+LostRuins@users.noreply.github.com> LostRuins Concedo <39025047+LostRuins@users.noreply.github.com> +Lucas Moura Belo Luciano Luo Tian Lyle Dean @@ -463,6 +487,7 @@ Matthew Tejo Matvey Soloviev Max Krasnyansky Max Krasnyansky +Maxim Evtush <154841002+maximevtush@users.noreply.github.com> Maxime <672982+maximegmd@users.noreply.github.com> Maximilian Winter Meng Zhang @@ -494,6 +519,7 @@ Miwa / Ensan <63481257+ensan-hcl@users.noreply.github.com> Mohammadreza Hendiani Mohammadreza Hendiani Molly Sophia +MoonRide303 <130458190+MoonRide303@users.noreply.github.com> MorganRO8 <47795945+MorganRO8@users.noreply.github.com> Murilo Santana Musab Gultekin @@ -524,6 +550,7 @@ Nikolas <127742645+nneubacher@users.noreply.github.com> Nindaleth Nuno OSecret <135510162+OLSecret@users.noreply.github.com> +Oleksandr Kuvshynov <661042+okuvshynov@users.noreply.github.com> Oleksandr Nikitin Oleksii Maryshchenko Olivier Chafik @@ -533,6 +560,7 @@ PAB Pablo Duboue Pascal Patry Patrice Ferlet +Patrick Peng Paul Tsochantaris Pavel Zloi Pavol Rusnak @@ -549,6 +577,7 @@ Pieter Ouwerkerk Plamen Minev Prashant Vithule <119530321+Vithulep@users.noreply.github.com> Przemysław Pawełczyk +PureJourney Qin Yue Chen <71813199+chenqiny@users.noreply.github.com> Qingyou Meng Qu Zongfu <43257352+yancaoweidaode@users.noreply.github.com> @@ -564,14 +593,17 @@ Rand Xie Randall Fitzgerald Random Fly Reinforce-II +Rémy O Rémy Oudompheng Ren Xuancheng Rene Leonhardt <65483435+reneleonhardt@users.noreply.github.com> Reza Kakhki +Reza Rahemtola <49811529+RezaRahemtola@users.noreply.github.com> RhinoDevel Riccardo Orlando Riceball LEE Rich Dougherty +Richard Richard Kiss Richard Roberson Rick G <26732651+TheFlipbook@users.noreply.github.com> @@ -588,6 +620,7 @@ Robert Sung-wook Shin Robey Holderith Robyn Roger Meier +Rohanjames1997 Roland <14355895+rbur0425@users.noreply.github.com> Romain Biessy Romain D <90720+Artefact2@users.noreply.github.com> @@ -610,6 +643,7 @@ Ryan Landay Ryder Wishart Ryuei Rőczey Barnabás <31726601+An0nie@users.noreply.github.com> +SAMI SRHMorris <69468379+SRHMorris@users.noreply.github.com> SXX SakuraUmi @@ -634,6 +668,8 @@ Shane A Shangning Xu <32517059+xushangning@users.noreply.github.com> Shankar Shanshan Shen <467638484@qq.com> +Shelby Jenkins <47464908+ShelbyJenkins@users.noreply.github.com> +Sheldon Robinson Shijie <821898965@qq.com> Shintarou Okada Shouzheng Liu <61452103+lshzh-ww@users.noreply.github.com> @@ -713,18 +749,24 @@ Victor Nogueira Victor Z. Peng Viet-Anh NGUYEN (Andrew) Vinesh Janarthanan <36610342+VJHack@users.noreply.github.com> +Vitali Lovich +Vivian Vlad Vladimir Vladimir Malyutin +Vladimir Vuksanovic <109677816+vvuksanovic@users.noreply.github.com> Vladimir Zorin VoidIsVoid <343750470@qq.com> Volodymyr Vitvitskyi <72226+signalpillar@users.noreply.github.com> +Wagner Bruna Wang Qin <37098874+wangqin0@users.noreply.github.com> Wang Ran (汪然) WangHaoranRobin <56047610+WangHaoranRobin@users.noreply.github.com> Weird Constructor +Weizhao Ouyang Welby Seely Wentai Zhang +Wilken Gottwalt <12194808+wgottwalt@users.noreply.github.com> WillCorticesAI <150854901+WillCorticesAI@users.noreply.github.com> William Tambellini William Tambellini @@ -816,6 +858,8 @@ chaihahaha chiranko <96988916+chiranko@users.noreply.github.com> clibdev <52199778+clibdev@users.noreply.github.com> clyang +cmdr2 +cmdr2 cocktailpeanut <121128867+cocktailpeanut@users.noreply.github.com> codezjx coezbek @@ -835,6 +879,7 @@ deepdiffuser <112834445+deepdiffuser@users.noreply.github.com> devojony <61173062+devojony@users.noreply.github.com> ditsuke divinity76 +dm4 dm4 dotpy314 <33351922+dotpy314@users.noreply.github.com> drbh @@ -849,6 +894,7 @@ fairydreaming <166155368+fairydreaming@users.noreply.github.com> fengerhu1 <2748250768@qq.com> fj-y-saito <85871716+fj-y-saito@users.noreply.github.com> fraxy-v <65565042+fraxy-v@users.noreply.github.com> +fxzjshm <11426482+fxzjshm@users.noreply.github.com> github-actions[bot] gliptic gn64 @@ -873,6 +919,7 @@ hydai iSma iacore <74560659+iacore@users.noreply.github.com> icppWorld <124377669+icppWorld@users.noreply.github.com> +igardev <49397134+igardev@users.noreply.github.com> igarnier intelmatt <61025942+intelmatt@users.noreply.github.com> iohub @@ -880,6 +927,7 @@ issixx <46835150+issixx@users.noreply.github.com> jacobi petrucciani <8117202+jpetrucciani@users.noreply.github.com> jaime-m-p <167997752+jaime-m-p@users.noreply.github.com> jameswu2014 <545426914@qq.com> +jason_w jdomke <28772296+jdomke@users.noreply.github.com> jiahao su jiez <373447296@qq.com> @@ -891,6 +939,7 @@ jon-chuang <9093549+jon-chuang@users.noreply.github.com> jp-x-g jukofyork <69222624+jukofyork@users.noreply.github.com> junchao-loongson <68935141+junchao-loongson@users.noreply.github.com> +junchao-zhao <68935141+junchao-loongson@users.noreply.github.com> jwj7140 <32943891+jwj7140@users.noreply.github.com> k.h.lai kaizau @@ -925,6 +974,7 @@ ltoniazzi <61414566+ltoniazzi@users.noreply.github.com> luoyu-intel m3ndax maddes8cht <55592906+maddes8cht@users.noreply.github.com> +magicse mahorozte <41834471+mahorozte@users.noreply.github.com> makomk manikbhandari @@ -935,6 +985,7 @@ matt23654 matteo mdrokz mgroeber9110 <45620825+mgroeber9110@users.noreply.github.com> +midnight minarchist mj-shifu <77107165+mj-shifu@users.noreply.github.com> mmyjona @@ -958,10 +1009,12 @@ omahs <73983677+omahs@users.noreply.github.com> oobabooga <112222186+oobabooga@users.noreply.github.com> opparco ostix360 <55257054+ostix360@users.noreply.github.com> +pascal-lc <49066376+pascal-lc@users.noreply.github.com> pculliton peidaqi pengxin99 perserk +petterreinholdtsen piDack <104877312+piDack@users.noreply.github.com> pmysl postmasters @@ -983,6 +1036,7 @@ semidark serhii-nakon <57632032+serhii-nakon@users.noreply.github.com> sharpHL <132747147+sharpHL@users.noreply.github.com> shibe2 +simon886212 <37953122+simon886212@users.noreply.github.com> singularity <12184989+singularity-s0@users.noreply.github.com> sjinzh sjxx <63994076+ylsdamxssjxxdd@users.noreply.github.com> @@ -1000,10 +1054,12 @@ tarcey tc-mb <157115220+tc-mb@users.noreply.github.com> texmex76 <40733439+texmex76@users.noreply.github.com> thement <40525767+thement@users.noreply.github.com> +theraininsky <76763719+theraininsky@users.noreply.github.com> thewh1teagle <61390950+thewh1teagle@users.noreply.github.com> tjohnman toyer <2042519524@qq.com> tslmy +tv1wnd <55383215+tv1wnd@users.noreply.github.com> ubik2 uint256_t uint256_t @@ -1014,6 +1070,7 @@ valiray <133289098+valiray@users.noreply.github.com> vb vik viric +vmobilis <75476228+vmobilis@users.noreply.github.com> vodkaslime <646329483@qq.com> vvhg1 <94630311+vvhg1@users.noreply.github.com> vxiiduu <73044267+vxiiduu@users.noreply.github.com> @@ -1028,6 +1085,8 @@ wzy <32936898+Freed-Wu@users.noreply.github.com> xaedes xaedes xctan +xiaobing318 <71554036+xiaobing318@users.noreply.github.com> +xiaofei xloem <0xloem@gmail.com> yangli2 ymcki <84055651+ymcki@users.noreply.github.com> diff --git a/Makefile b/Makefile index 5339d490b4..1f9455eff0 100644 --- a/Makefile +++ b/Makefile @@ -836,7 +836,7 @@ ifdef GGML_MUSA else MUSA_PATH ?= /opt/musa endif - MUSA_ARCHITECTURES ?= 21;22 + MUSA_ARCHITECTURES ?= 21;22;31 MK_CPPFLAGS += -DGGML_USE_MUSA -DGGML_USE_CUDA MK_LDFLAGS += -L$(MUSA_PATH)/lib -Wl,-rpath=$(MUSA_PATH)/lib diff --git a/README.md b/README.md index d73b0495d8..1eec944f27 100644 --- a/README.md +++ b/README.md @@ -157,6 +157,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo - Guile Scheme: [guile_llama_cpp](https://savannah.nongnu.org/projects/guile-llama-cpp) - Swift [srgtuszy/llama-cpp-swift](https://github.com/srgtuszy/llama-cpp-swift) - Swift [ShenghaiWang/SwiftLlama](https://github.com/ShenghaiWang/SwiftLlama) +- Delphi [Embarcadero/llama-cpp-delphi](https://github.com/Embarcadero/llama-cpp-delphi) @@ -171,6 +172,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo - [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT) - [iohub/collama](https://github.com/iohub/coLLaMA) (Apache-2.0) - [janhq/jan](https://github.com/janhq/jan) (AGPL) +- [johnbean393/Sidekick](https://github.com/johnbean393/Sidekick) (MIT) - [KanTV](https://github.com/zhouwg/kantv?tab=readme-ov-file) (Apache-2.0) - [KodiBot](https://github.com/firatkiral/kodibot) (GPL) - [llama.vim](https://github.com/ggml-org/llama.vim) (MIT) diff --git a/ci/run.sh b/ci/run.sh index 77c32ce005..9fc19c89d8 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -352,10 +352,10 @@ function gg_run_open_llama_7b_v2 { (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log - (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 10 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log - (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log - (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log - (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log function check_ppl { qnt="$1" diff --git a/common/arg.cpp b/common/arg.cpp index 3e549ede0e..fe6a1eece7 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -764,7 +764,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_env("LLAMA_ARG_CTX_SIZE")); add_opt(common_arg( {"-n", "--predict", "--n-predict"}, "N", - string_format("number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)", params.n_predict), + string_format( + ex == LLAMA_EXAMPLE_MAIN || ex == LLAMA_EXAMPLE_INFILL + ? "number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)" + : "number of tokens to predict (default: %d, -1 = infinity)", + params.n_predict), [](common_params & params, int value) { params.n_predict = value; } @@ -1867,16 +1871,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_examples({LLAMA_EXAMPLE_PASSKEY})); add_opt(common_arg( {"-o", "--output", "--output-file"}, "FNAME", - string_format("output file (default: '%s')", - ex == LLAMA_EXAMPLE_EXPORT_LORA - ? params.lora_outfile.c_str() - : ex == LLAMA_EXAMPLE_CVECTOR_GENERATOR - ? params.cvector_outfile.c_str() - : params.out_file.c_str()), + string_format("output file (default: '%s')", params.out_file.c_str()), [](common_params & params, const std::string & value) { params.out_file = value; - params.cvector_outfile = value; - params.lora_outfile = value; } ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA})); add_opt(common_arg( @@ -2571,5 +2568,43 @@ common_params_context common_params_parser_init(common_params & params, llama_ex } ).set_examples({LLAMA_EXAMPLE_SERVER})); + add_opt(common_arg( + {"--fim-qwen-7b-spec"}, + string_format("use Qwen 2.5 Coder 7B + 0.5B draft for speculative decoding (note: can download weights from the internet)"), + [](common_params & params) { + params.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF"; + params.hf_file = "qwen2.5-coder-7b-q8_0.gguf"; + params.speculative.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF"; + params.speculative.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf"; + params.speculative.n_gpu_layers = 99; + params.port = 8012; + params.n_gpu_layers = 99; + params.flash_attn = true; + params.n_ubatch = 1024; + params.n_batch = 1024; + params.n_ctx = 0; + params.n_cache_reuse = 256; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + + add_opt(common_arg( + {"--fim-qwen-14b-spec"}, + string_format("use Qwen 2.5 Coder 14B + 0.5B draft for speculative decoding (note: can download weights from the internet)"), + [](common_params & params) { + params.hf_repo = "ggml-org/Qwen2.5-Coder-14B-Q8_0-GGUF"; + params.hf_file = "qwen2.5-coder-14b-q8_0.gguf"; + params.speculative.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF"; + params.speculative.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf"; + params.speculative.n_gpu_layers = 99; + params.port = 8012; + params.n_gpu_layers = 99; + params.flash_attn = true; + params.n_ubatch = 1024; + params.n_batch = 1024; + params.n_ctx = 0; + params.n_cache_reuse = 256; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + return ctx_arg; } diff --git a/common/chat.cpp b/common/chat.cpp index 1b10219cca..62ca26ad76 100644 --- a/common/chat.cpp +++ b/common/chat.cpp @@ -60,7 +60,9 @@ std::vector common_chat_msgs_parse_oaicompat(const json & messa } msg.role = message.at("role"); - if (message.contains("content")) { + auto has_content = message.contains("content"); + auto has_tool_calls = message.contains("tool_calls"); + if (has_content) { const auto & content = message.at("content"); if (content.is_string()) { msg.content = content; @@ -81,19 +83,8 @@ std::vector common_chat_msgs_parse_oaicompat(const json & messa } else if (!content.is_null()) { throw std::runtime_error("Invalid 'content' type: expected string or array, got " + content.dump() + " (ref: https://github.com/ggml-org/llama.cpp/issues/8367)"); } - } else { - throw std::runtime_error("Expected 'content' (ref: https://github.com/ggml-org/llama.cpp/issues/8367)"); } - if (message.contains("reasoning_content")) { - msg.reasoning_content = message.at("reasoning_content"); - } - if (message.contains("name")) { - msg.tool_name = message.at("name"); - } - if (message.contains("tool_call_id")) { - msg.tool_call_id = message.at("tool_call_id"); - } - if (message.contains("tool_calls")) { + if (has_tool_calls) { for (const auto & tool_call : message.at("tool_calls")) { common_chat_tool_call tc; if (!tool_call.contains("type")) { @@ -118,6 +109,18 @@ std::vector common_chat_msgs_parse_oaicompat(const json & messa msg.tool_calls.push_back(tc); } } + if (!has_content && !has_tool_calls) { + throw std::runtime_error("Expected 'content' or 'tool_calls' (ref: https://github.com/ggml-org/llama.cpp/issues/8367 & https://github.com/ggml-org/llama.cpp/issues/12279)"); + } + if (message.contains("reasoning_content")) { + msg.reasoning_content = message.at("reasoning_content"); + } + if (message.contains("name")) { + msg.tool_name = message.at("name"); + } + if (message.contains("tool_call_id")) { + msg.tool_call_id = message.at("tool_call_id"); + } msgs.push_back(msg); } @@ -442,6 +445,7 @@ std::string common_chat_format_name(common_chat_format format) { case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2: return "Functionary v3.2"; case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1: return "Functionary v3.1 Llama 3.1"; case COMMON_CHAT_FORMAT_HERMES_2_PRO: return "Hermes 2 Pro"; + case COMMON_CHAT_FORMAT_HERMES_2_PRO_EXTRACT_REASONING: return "Hermes 2 Pro (extract reasoning)"; case COMMON_CHAT_FORMAT_COMMAND_R7B: return "Command R7B"; case COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING: return "Command R7B (extract reasoning)"; default: @@ -875,9 +879,9 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_ return data; } static common_chat_msg common_chat_parse_command_r7b(const std::string & input, bool extract_reasoning) { - static std::regex thought_regex("(<\\|START_THINKING\\|>([\\s\\S]*?)<\\|END_THINKING\\|>)([\\s\\S]*)"); - static std::regex action_regex("<\\|START_ACTION\\|>([\\s\\S]*?)<\\|END_ACTION\\|>"); - static std::regex response_regex("(?:<\\|START_RESPONSE\\|>)?([\\s\\S]*?)<\\|END_RESPONSE\\|>"); + static const std::regex thought_regex("(<\\|START_THINKING\\|>([\\s\\S]*?)<\\|END_THINKING\\|>)([\\s\\S]*)"); + static const std::regex action_regex("<\\|START_ACTION\\|>([\\s\\S]*?)<\\|END_ACTION\\|>"); + static const std::regex response_regex("(?:<\\|START_RESPONSE\\|>)?([\\s\\S]*?)<\\|END_RESPONSE\\|>"); std::smatch match; @@ -1009,10 +1013,10 @@ static common_chat_params common_chat_params_init_llama_3_1_tool_calls(const com } static common_chat_msg common_chat_parse_llama_3_1(const std::string & input, bool with_builtin_tools = false) { // TODO: tighten & simplify the parser, don't accept leading text context. - static std::regex function_regex( + static const std::regex function_regex( "\\s*\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"([^\"]+)\"\\s*,\\s*\"parameters\"\\s*: "); - static std::regex close_regex("\\}\\s*"); - static std::regex builtin_call_regex("<\\|python_tag\\|>\\s*([^.(]+)\\s*\\.\\s*call\\s*\\(\\s*([\\w]+)\\s*=\\s*([\\s\\S]*?)\\)"); + static const std::regex close_regex("\\}\\s*"); + static const std::regex builtin_call_regex("<\\|python_tag\\|>\\s*([^.(]+)\\s*\\.\\s*call\\s*\\(\\s*([\\w]+)\\s*=\\s*([\\s\\S]*?)\\)"); if (with_builtin_tools) { std::smatch match; @@ -1102,34 +1106,42 @@ static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_ data.format = inputs.extract_reasoning ? COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING : COMMON_CHAT_FORMAT_DEEPSEEK_R1; return data; } -static common_chat_msg common_chat_parse_deepseek_r1(const std::string & input, bool extract_reasoning) { - static std::regex function_regex("<｜tool▁call▁begin｜>function<｜tool▁sep｜>([^\n]+)\n```json\n"); - static std::regex close_regex("```[\\s\\r\\n]*<｜tool▁call▁end｜>"); - static std::regex reasoning_content_regex("((?:)?([\\s\\S\\r\\n]*?))?([\\s\\S\\r\\n]*)"); - static std::regex tool_calls_regex("[\\s\\r\\n]*(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>)([\\s\\S\\r\\n]*?)<｜tool▁calls▁end｜>"); - common_chat_msg msg; - msg.role = "assistant"; +static common_chat_msg handle_think_tag_prelude(const std::string & input, bool extract_reasoning, const std::function & rest_parser) { std::smatch match; + static const std::regex reasoning_content_regex("((?:)?([\\s\\S\\r\\n]*?))?([\\s\\S\\r\\n]*)"); if (std::regex_match(input, match, reasoning_content_regex)) { - std::string rest; + auto rest = match[3].str(); + auto msg = rest_parser(rest); + auto reasoning_content = string_strip(match[2].str()); if (extract_reasoning) { - msg.reasoning_content = string_strip(match[2].str()); - } else { - msg.content = match[1].str(); + msg.reasoning_content = reasoning_content; + } else if (!reasoning_content.empty()) { + std::ostringstream content; + content << "" << reasoning_content << "" << msg.content; + msg.content = content.str(); } - rest = match[3].str(); + return msg; + } + return rest_parser(input); +} +static common_chat_msg common_chat_parse_deepseek_r1(const std::string & input, bool extract_reasoning) { + return handle_think_tag_prelude(input, extract_reasoning, [](const std::string & input) { + static const std::regex function_regex("<｜tool▁call▁begin｜>function<｜tool▁sep｜>([^\n]+)\n```json\n"); + static const std::regex close_regex("```[\\s\\r\\n]*<｜tool▁call▁end｜>"); + static const std::regex tool_calls_regex("[\\s\\r\\n]*(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>)([\\s\\S\\r\\n]*?)<｜tool▁calls▁end｜>"); - if (std::regex_search(rest, match, tool_calls_regex)) { + common_chat_msg msg; + msg.role = "assistant"; + std::smatch match; + if (std::regex_search(input, match, tool_calls_regex)) { auto tool_calls = match[1].str(); auto msg2 = parse_json_tool_calls(tool_calls, std::nullopt, function_regex, close_regex); msg.tool_calls = std::move(msg2.tool_calls); } else { - msg.content += std::string(rest.begin() + rest.find_first_not_of(" \r\n"), rest.end()); + msg.content = input; } - } else { - msg.content = input; - } - return msg; + return msg; + }); } static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct templates_params & inputs) { @@ -1234,8 +1246,8 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_ } static common_chat_msg common_chat_parse_functionary_v3_2(const std::string & input) { - static std::regex function_regex(R"((?:>>>)?(?:assistant<|end_header_id|>\n)?(\w+)\n)"); - static std::regex close_regex(R"($|(?=>>>))"); + static const std::regex function_regex(R"((?:>>>)?(?:assistant<|end_header_id|>\n)?(\w+)\n)"); + static const std::regex close_regex(R"($|(?=>>>))"); std::string content; auto it = input.begin(); @@ -1324,7 +1336,7 @@ static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(con } static common_chat_msg common_chat_parse_functionary_v3_1_llama_3_1(const std::string & input) { // This version of Functionary still supports the llama 3.1 tool call format for the python tool. - static std::regex python_tag_regex(R"(<\|python_tag\|>([\s\S\n]*)$)"); + static const std::regex python_tag_regex(R"(<\|python_tag\|>([\s\S\n]*)$)"); std::smatch match; if (std::regex_search(input, match, python_tag_regex)) { auto code = match[1].str(); @@ -1338,8 +1350,8 @@ static common_chat_msg common_chat_parse_functionary_v3_1_llama_3_1(const std::s }); return msg; } - static std::regex function_regex(R"()"); - static std::regex close_regex(R"()"); + static const std::regex function_regex(R"()"); + static const std::regex close_regex(R"()"); // TODO: tighten & simplify. return parse_json_tool_calls(input, std::nullopt, function_regex, close_regex); } @@ -1406,6 +1418,8 @@ static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat "(?:```(?:json|xml)?\n\\s*)?(?:|||)?\\s*\\{\\s*\"", //name\"\\s*:\\s*\"" + escaped_name + "\"", }); data.preserved_tokens = { + "", + "", "", "", "" // match 2 (open_tag) - "|" - "|" - "|" - "|" - "|" - "|" - "|" - ")?" - "(\\s*\\{\\s*\"name\"\\s*:[\\s\\S]*)" // match 3 (named tool call + rest) - ")" - "|" - "(?:]+)>" // match 4 (function name) - "|)" // match 5 (function name again) - "([\\s\\S]*)" // match 6 (function arguments + rest)})" - ); +static common_chat_msg common_chat_parse_hermes_2_pro(const std::string& input, bool extract_reasoning) { + return handle_think_tag_prelude(input, extract_reasoning, [](const std::string & input) { + static const std::regex open_regex( + "(?:" + "(```(?:xml|json)?\\n\\s*)?" // match 1 (block_start) + "(" // match 2 (open_tag) + "|" + "|" + "|" + "|" + "|" + "|" + "|" + ")?" + "(\\s*\\{\\s*\"name\"\\s*:[\\s\\S]*)" // match 3 (named tool call + rest) + ")" + "|" + "(?:]+)>" // match 4 (function name) + "|)" // match 5 (function name again) + "([\\s\\S]*)" // match 6 (function arguments + rest)})" + ); - try { + try { + common_chat_msg msg; + msg.role = "assistant"; - common_chat_msg msg; - msg.role = "assistant"; + std::string::const_iterator it = input.begin(); + const std::string::const_iterator end = input.end(); + std::smatch match; - std::string::const_iterator it = input.begin(); - const std::string::const_iterator end = input.end(); - std::smatch match; + while (it != end) { + if (std::regex_search(it, end, match, open_regex)) { + // Add content before the match + msg.content += std::string(it, match[0].first); - while (it != end) { - if (std::regex_search(it, end, match, open_regex)) { - // Add content before the match - msg.content += std::string(it, match[0].first); + auto block_start = match[1].str(); + std::string block_end = block_start.empty() ? "" : "```"; - auto block_start = match[1].str(); - std::string block_end = block_start.empty() ? "" : "```"; + auto open_tag = match[2].str(); + std::string close_tag; - auto open_tag = match[2].str(); - std::string close_tag; + if (match[3].matched) { + close_tag = open_tag.empty() ? "" : ""; + // Start parsing from after the opening tags + auto json_it = match[6].first; + json arguments; + if (parse_json(json_it, end, arguments)) { + msg.tool_calls.emplace_back(process_tool_call({ + {"name", function_name}, + {"arguments", arguments}, + })); + it = json_it; // Move iterator past parsed JSON + + // Handle close tags + consume_spaces(it, end); + if (!close_tag.empty() && !parse_literal(it, end, close_tag)) { + throw std::runtime_error("Failed to parse closing tag"); + } + consume_spaces(it, end); + if (!block_end.empty() && !parse_literal(it, end, block_end)) { + throw std::runtime_error("Failed to parse block end"); + } + consume_spaces(it, end); + } else { + // Not a valid tool call, treat as content + msg.content += std::string(match[0].first, match[0].second); + it = match[0].second; + } } } else { - auto function_name = match[4].str(); - if (function_name.empty()) { - function_name = match[5].str(); - } - GGML_ASSERT(!function_name.empty()); - - close_tag = ""; - // Start parsing from after the opening tags - auto json_it = match[6].first; - json arguments; - if (parse_json(json_it, end, arguments)) { - msg.tool_calls.emplace_back(process_tool_call({ - {"name", function_name}, - {"arguments", arguments}, - })); - it = json_it; // Move iterator past parsed JSON - - // Handle close tags - consume_spaces(it, end); - if (!close_tag.empty() && !parse_literal(it, end, close_tag)) { - throw std::runtime_error("Failed to parse closing tag"); - } - consume_spaces(it, end); - if (!block_end.empty() && !parse_literal(it, end, block_end)) { - throw std::runtime_error("Failed to parse block end"); - } - consume_spaces(it, end); - } else { - // Not a valid tool call, treat as content - msg.content += std::string(match[0].first, match[0].second); - it = match[0].second; - } + // Add remaining content + msg.content += std::string(it, end); + break; } - } else { - // Add remaining content - msg.content += std::string(it, end); - break; } + return msg; + } catch (const std::exception & e) { + LOG_ERR("Failed to parse hermes 2 pro input: %s\n", e.what()); + common_chat_msg msg; + msg.role = "assistant"; + msg.content = input; + return msg; } - return msg; - } catch (const std::exception & e) { - LOG_ERR("Failed to parse hermes 2 pro input: %s\n", e.what()); - common_chat_msg msg; - msg.role = "assistant"; - msg.content = input; - return msg; - } + }); } static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) { @@ -1606,6 +1621,11 @@ static common_chat_params common_chat_templates_apply_jinja( return common_chat_params_init_command_r7b(tmpl, params); } + // Hermes 2/3 Pro, Qwen 2.5 Instruct (w/ tools) + if (src.find("") != std::string::npos && params.json_schema.is_null()) { + return common_chat_params_init_hermes_2_pro(tmpl, params); + } + // Use generic handler when mixing tools + JSON schema. // TODO: support that mix in handlers below. if ((params.tools.is_array() && params.json_schema.is_object())) { @@ -1627,11 +1647,6 @@ static common_chat_params common_chat_templates_apply_jinja( return common_chat_params_init_without_tools(tmpl, params); } - // Hermes 2/3 Pro, Qwen 2.5 Instruct (w/ tools) - if (src.find("") != std::string::npos) { - return common_chat_params_init_hermes_2_pro(tmpl, params); - } - // Functionary v3.1 (w/ tools) if (src.find("<|start_header_id|>") != std::string::npos && src.find(" split(const std::string & s, const std::string & sep) { + std::vector result; + size_t start = 0; + size_t end = s.find(sep); + while (end != std::string::npos) { + result.push_back(s.substr(start, end - start)); + start = end + sep.length(); + end = s.find(sep, start); + } + result.push_back(s.substr(start)); + return result; +} + static std::string capitalize(const std::string & s) { if (s.empty()) return s; auto result = s; @@ -1467,8 +1481,26 @@ public: } else if (obj.is_string()) { auto str = obj.get(); if (method->get_name() == "strip") { - vargs.expectArgs("strip method", {0, 0}, {0, 0}); - return Value(strip(str)); + vargs.expectArgs("strip method", {0, 1}, {0, 0}); + auto chars = vargs.args.empty() ? "" : vargs.args[0].get(); + return Value(strip(str, chars)); + } else if (method->get_name() == "lstrip") { + vargs.expectArgs("lstrip method", {0, 1}, {0, 0}); + auto chars = vargs.args.empty() ? "" : vargs.args[0].get(); + return Value(strip(str, chars, /* left= */ true, /* right= */ false)); + } else if (method->get_name() == "rstrip") { + vargs.expectArgs("rstrip method", {0, 1}, {0, 0}); + auto chars = vargs.args.empty() ? "" : vargs.args[0].get(); + return Value(strip(str, chars, /* left= */ false, /* right= */ true)); + } else if (method->get_name() == "split") { + vargs.expectArgs("split method", {1, 1}, {0, 0}); + auto sep = vargs.args[0].get(); + auto parts = split(str, sep); + Value result = Value::array(); + for (const auto& part : parts) { + result.push_back(Value(part)); + } + return result; } else if (method->get_name() == "capitalize") { vargs.expectArgs("capitalize method", {0, 0}, {0, 0}); return Value(capitalize(str)); diff --git a/common/speculative.cpp b/common/speculative.cpp index 1bac3a1ce1..ccad70fa9e 100644 --- a/common/speculative.cpp +++ b/common/speculative.cpp @@ -173,7 +173,7 @@ llama_tokens common_speculative_gen_draft( result.reserve(params.n_draft); if (reuse_n == 0) { - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); prompt.clear(); } else { @@ -192,14 +192,14 @@ llama_tokens common_speculative_gen_draft( } if (reuse_i > 0) { - llama_kv_cache_seq_rm (ctx, 0, 0, reuse_i); - llama_kv_cache_seq_add(ctx, 0, reuse_i, -1, -reuse_i); + llama_kv_self_seq_rm (ctx, 0, 0, reuse_i); + llama_kv_self_seq_add(ctx, 0, reuse_i, -1, -reuse_i); prompt.erase(prompt.begin(), prompt.begin() + reuse_i); } if (reuse_n < (int) prompt.size()) { - llama_kv_cache_seq_rm (ctx, 0, reuse_n, -1); + llama_kv_self_seq_rm (ctx, 0, reuse_n, -1); prompt.erase(prompt.begin() + reuse_n, prompt.end()); } diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 6358a94e9b..b5d95bd563 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -861,6 +861,9 @@ class Model: for token_id, token_data in added_tokens_decoder.items(): token_id = int(token_id) token: str = token_data["content"] + if token_id >= vocab_size: + logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}') + continue if toktypes[token_id] != SentencePieceTokenTypes.UNUSED: if tokens[token_id] != token.encode("utf-8"): logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token!r}') @@ -3322,6 +3325,83 @@ class Gemma2Model(Model): return [(self.map_tensor_name(name), data_torch)] +@Model.register("Gemma3ForCausalLM", "Gemma3ForConditionalGeneration") +class Gemma3Model(Model): + model_arch = gguf.MODEL_ARCH.GEMMA3 + has_vision: bool = False + + # we need to merge the text_config into the root level of hparams + def __init__(self, *args, **kwargs): + hparams = Model.load_hparams(kwargs["dir_model"]) + if "text_config" in hparams: + hparams = {**hparams, **hparams["text_config"]} + kwargs["hparams"] = hparams + super().__init__(*args, **kwargs) + if "vision_config" in hparams: + logger.info("Has vision encoder, but it will be ignored") + self.has_vision = True + + def write(self): + super().write() + if self.has_vision: + logger.info("NOTE: this script only convert the language model to GGUF") + logger.info(" for the vision model, please use gemma3_convert_encoder_to_gguf.py") + + def set_vocab(self): + self._set_vocab_sentencepiece() + + self.gguf_writer.add_add_space_prefix(False) + + def set_gguf_parameters(self): + hparams = self.hparams + block_count = hparams["num_hidden_layers"] + + # some default values are not specified in the hparams + self.gguf_writer.add_context_length(hparams.get("max_position_embeddings", 131072)) + self.gguf_writer.add_embedding_length(hparams["hidden_size"]) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"]) + self.gguf_writer.add_head_count(hparams.get("num_attention_heads", 8)) + self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("rms_norm_eps", 1e-6)) + self.gguf_writer.add_key_length(hparams.get("head_dim", 256)) + self.gguf_writer.add_value_length(hparams.get("head_dim", 256)) + self.gguf_writer.add_file_type(self.ftype) + self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 1_000_000.0)) # for global layers + # both attn_logit_softcapping and final_logit_softcapping are removed in Gemma3 + assert hparams.get("attn_logit_softcapping") is None + assert hparams.get("final_logit_softcapping") is None + self.gguf_writer.add_sliding_window(hparams["sliding_window"]) + self.gguf_writer.add_head_count_kv(hparams.get("num_key_value_heads", 4)) + if hparams.get("rope_scaling") is not None: + assert hparams["rope_scaling"]["rope_type"] == "linear" + # important: this rope_scaling is only applied for global layers, and not used by 1B model + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"]) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + del bid # unused + + if name.startswith("language_model."): + name = name.replace("language_model.", "") + elif name.startswith("multi_modal_projector.") or name.startswith("vision_tower.") \ + or name.startswith("multimodal_projector.") or name.startswith("vision_model."): # this is for old HF model, should be removed later + # ignore vision tensors + return [] + + # remove OOV (out-of-vocabulary) rows in token_embd + if "embed_tokens.weight" in name: + vocab = self._create_vocab_sentencepiece() + tokens = vocab[0] + data_torch = data_torch[:len(tokens)] + + # ref code in Gemma3RMSNorm + # output = output * (1.0 + self.weight.float()) + if name.endswith("norm.weight"): + data_torch = data_torch + 1 + + return [(self.map_tensor_name(name), data_torch)] + + @Model.register("Starcoder2ForCausalLM") class StarCoder2Model(Model): model_arch = gguf.MODEL_ARCH.STARCODER2 diff --git a/docs/build.md b/docs/build.md index b3ecf043d7..2e3975c145 100644 --- a/docs/build.md +++ b/docs/build.md @@ -197,29 +197,53 @@ The following compilation options are also available to tweak performance: ## MUSA -This provides GPU acceleration using the MUSA cores of your Moore Threads MTT GPU. Make sure to have the MUSA SDK installed. You can download it from here: [MUSA SDK](https://developer.mthreads.com/sdk/download/musa). +This provides GPU acceleration using a Moore Threads GPU. Make sure to have the [MUSA SDK](https://developer.mthreads.com/musa/musa-sdk) installed. -- Using `CMake`: +#### Download directly from Moore Threads - ```bash - cmake -B build -DGGML_MUSA=ON - cmake --build build --config Release +You may find the official downloads here: [Moore Threads developer site](https://developer.mthreads.com/sdk/download/musa). + +### Compilation + +```bash +cmake -B build -DGGML_MUSA=ON +cmake --build build --config Release +``` + +#### Override Compute Capability Specifications + +By default, all supported compute capabilities are enabled. To customize this behavior, you can specify the `MUSA_ARCHITECTURES` option in the CMake command: + +```bash +cmake -B build -DGGML_MUSA=ON -DMUSA_ARCHITECTURES="21" +``` + +This configuration enables only compute capability `2.1` (MTT S80) during compilation, which can help reduce compilation time. + +#### Compilation options + +Most of the compilation options available for CUDA should also be available for MUSA, though they haven't been thoroughly tested yet. + +- For static builds, add `-DBUILD_SHARED_LIBS=OFF` and `-DCMAKE_POSITION_INDEPENDENT_CODE=ON`: ``` - - For static build: - - ```bash cmake -B build -DGGML_MUSA=ON \ -DBUILD_SHARED_LIBS=OFF -DCMAKE_POSITION_INDEPENDENT_CODE=ON cmake --build build --config Release ``` -The environment variable [`MUSA_VISIBLE_DEVICES`](https://docs.mthreads.com/musa-sdk/musa-sdk-doc-online/programming_guide/Z%E9%99%84%E5%BD%95/) can be used to specify which GPU(s) will be used. +### Runtime MUSA environmental variables + +You may set the [musa environmental variables](https://docs.mthreads.com/musa-sdk/musa-sdk-doc-online/programming_guide/Z%E9%99%84%E5%BD%95/) at runtime. + +```bash +# Use `MUSA_VISIBLE_DEVICES` to hide the first compute device. +MUSA_VISIBLE_DEVICES="-0" ./build/bin/llama-server --model /srv/models/llama.gguf +``` + +### Unified Memory The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. -Most of the compilation options available for CUDA should also be available for MUSA, though they haven't been thoroughly tested yet. - ## HIP This provides GPU acceleration on HIP-supported AMD GPUs. @@ -235,6 +259,12 @@ You can download it from your Linux distro's package manager or from here: [ROCm On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`. However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs). + To enhance flash attention performance on RDNA3+ or CDNA architectures, you can utilize the rocWMMA library by enabling the `-DGGML_HIP_ROCWMMA_FATTN=ON` option. This requires rocWMMA headers to be installed on the build system. + + The rocWMMA library is included by default when installing the ROCm SDK using the `rocm` meta package provided by AMD. Alternatively, if you are not using the meta package, you can install the library using the `rocwmma-dev` or `rocwmma-devel` package, depending on your system's package manager. + + As an alternative, you can manually install the library by cloning it from the official [GitHub repository](https://github.com/ROCm/rocWMMA), checkout the corresponding version tag (e.g. `rocm-6.2.4`) and set `-DCMAKE_CXX_FLAGS="-I/library/include/"` in CMake. This also works under Windows despite not officially supported by AMD. + Note that if you get the following error: ``` clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library diff --git a/docs/function-calling.md b/docs/function-calling.md index 92cb6531ab..c3873c3fa6 100644 --- a/docs/function-calling.md +++ b/docs/function-calling.md @@ -287,30 +287,32 @@ Here are some models known to work (w/ chat template override when needed): llama-server --jinja -fa -hf bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M llama-server --jinja -fa -hf bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q6_K_L -llama-server --jinja -fa -hf bartowski/functionary-small-v3.2-GGUF:Q4_K_M llama-server --jinja -fa -hf bartowski/Llama-3.3-70B-Instruct-GGUF:Q4_K_M -# Native support for DeepSeek R1 works best w/ our own template (official template buggy) +# Native support for DeepSeek R1 works best w/ our template override (official template is buggy, although we do work around it) llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q6_K_L \ ---chat-template-file models/templates/llama-cpp-deepseek-r1.jinja + --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M \ ---chat-template-file models/templates/llama-cpp-deepseek-r1.jinja + --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja # Native support requires the right template for these GGUFs: +llama-server --jinja -fa -hf bartowski/functionary-small-v3.2-GGUF:Q4_K_M + --chat-template-file models/templates/meetkai-functionary-medium-v3.2.jinja + llama-server --jinja -fa -hf bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M \ ---chat-template-file <( python scripts/get_chat_template.py NousResearch/Hermes-2-Pro-Llama-3-8B tool_use ) + --chat-template-file models/templates/NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja llama-server --jinja -fa -hf bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M \ ---chat-template-file <( python scripts/get_chat_template.py NousResearch/Hermes-3-Llama-3.1-8B tool_use ) + --chat-template-file models/templates/NousResearch-Hermes-3-Llama-3.1-8B-tool_use.jinja llama-server --jinja -fa -hf bartowski/firefunction-v2-GGUF -hff firefunction-v2-IQ1_M.gguf \ ---chat-template-file <( python scripts/get_chat_template.py fireworks-ai/llama-3-firefunction-v2 tool_use ) + --chat-template-file models/templates/fireworks-ai-llama-3-firefunction-v2.jinja llama-server --jinja -fa -hf bartowski/c4ai-command-r7b-12-2024-GGUF:Q6_K_L \ ---chat-template-file <( python scripts/get_chat_template.py CohereForAI/c4ai-command-r7b-12-2024 tool_use ) + --chat-template-file models/templates/CohereForAI-c4ai-command-r7b-12-2024-tool_use.jinja # Generic format support llama-server --jinja -fa -hf bartowski/phi-4-GGUF:Q4_0 @@ -318,6 +320,8 @@ llama-server --jinja -fa -hf bartowski/gemma-2-2b-it-GGUF:Q8_0 llama-server --jinja -fa -hf bartowski/c4ai-command-r-v01-GGUF:Q2_K ``` +To get the official template from original HuggingFace repos, you can use [scripts/get_chat_template.py](../scripts/get_chat_template.py) (see examples invocations in [models/templates/README.md](../models/templates/README.md)) + > [!TIP] > If there is no official `tool_use` Jinja template, you may want to set `--chat-template chatml` to use a default that works with many models (YMMV!), or write your own (e.g. we provide a custom [llama-cpp-deepseek-r1.jinja](../models/templates/llama-cpp-deepseek-r1.jinja) for DeepSeek R1 distills) diff --git a/examples/batched-bench/batched-bench.cpp b/examples/batched-bench/batched-bench.cpp index 0659ab6f11..430e8be512 100644 --- a/examples/batched-bench/batched-bench.cpp +++ b/examples/batched-bench/batched-bench.cpp @@ -132,7 +132,7 @@ int main(int argc, char ** argv) { const auto t_pp_start = ggml_time_us(); - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); if (!decode_helper(ctx, batch, ctx_params.n_batch)) { LOG_ERR("%s: llama_decode() failed\n", __func__); @@ -141,7 +141,7 @@ int main(int argc, char ** argv) { if (is_pp_shared) { for (int32_t i = 1; i < pl; ++i) { - llama_kv_cache_seq_cp(ctx, 0, i, -1, -1); + llama_kv_self_seq_cp(ctx, 0, i, -1, -1); } } diff --git a/examples/batched.swift/Sources/main.swift b/examples/batched.swift/Sources/main.swift index 55c31166ca..514989e340 100644 --- a/examples/batched.swift/Sources/main.swift +++ b/examples/batched.swift/Sources/main.swift @@ -116,7 +116,7 @@ if llama_decode(context, batch) != 0 { } for i in 1 ..< n_parallel { - llama_kv_cache_seq_cp(context, 0, Int32(i), 0, batch.n_tokens) + llama_kv_self_seq_cp(context, 0, Int32(i), 0, batch.n_tokens) } if n_parallel > 1 { diff --git a/examples/cvector-generator/cvector-generator.cpp b/examples/cvector-generator/cvector-generator.cpp index 413b71d34c..2a90715501 100644 --- a/examples/cvector-generator/cvector-generator.cpp +++ b/examples/cvector-generator/cvector-generator.cpp @@ -342,7 +342,7 @@ static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) { } static bool get_hidden_layers(llama_context * ctx, std::vector & tokens) { - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) { fprintf(stderr, "%s : failed to eval\n", __func__); return false; @@ -394,6 +394,8 @@ static int prepare_entries(common_params & params, train_context & ctx_train) { int main(int argc, char ** argv) { common_params params; + params.out_file = "control_vector.gguf"; + if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage)) { return 1; } @@ -498,7 +500,7 @@ int main(int argc, char ** argv) { } // write output vectors to gguf - export_gguf(ctx_train.v_final, params.cvector_outfile, model_hint); + export_gguf(ctx_train.v_final, params.out_file, model_hint); llama_backend_free(); diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index 3dd9f2b07d..6f08904159 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -38,7 +38,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu const struct llama_model * model = llama_get_model(ctx); // clear previous kv_cache values (irrelevant for embeddings) - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); // run model LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq); diff --git a/examples/export-lora/export-lora.cpp b/examples/export-lora/export-lora.cpp index 91238e4beb..e7d0fbfffe 100644 --- a/examples/export-lora/export-lora.cpp +++ b/examples/export-lora/export-lora.cpp @@ -413,20 +413,22 @@ static void print_usage(int, char ** argv) { int main(int argc, char ** argv) { common_params params; + params.out_file = "ggml-lora-merged-f16.gguf"; + if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage)) { return 1; } g_verbose = (params.verbosity > 1); try { - lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.cpuparams.n_threads); + lora_merge_ctx ctx(params.model, params.lora_adapters, params.out_file, params.cpuparams.n_threads); ctx.run_merge(); } catch (const std::exception & err) { fprintf(stderr, "%s\n", err.what()); exit(EXIT_FAILURE); } - printf("done, output file is %s\n", params.lora_outfile.c_str()); + printf("done, output file is %s\n", params.out_file.c_str()); return 0; } diff --git a/examples/gritlm/gritlm.cpp b/examples/gritlm/gritlm.cpp index 72eb462574..f7db7861c1 100644 --- a/examples/gritlm/gritlm.cpp +++ b/examples/gritlm/gritlm.cpp @@ -45,7 +45,7 @@ static std::vector> encode(llama_context * ctx, const std::ve } // clear previous kv_cache values (irrelevant for embeddings) - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); llama_set_embeddings(ctx, true); llama_set_causal_attn(ctx, false); @@ -102,7 +102,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std llama_token eos_token = llama_vocab_eos(vocab); - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); llama_set_embeddings(ctx, false); llama_set_causal_attn(ctx, true); diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp index 4edc0bfacf..31b675e8f9 100644 --- a/examples/imatrix/imatrix.cpp +++ b/examples/imatrix/imatrix.cpp @@ -206,9 +206,6 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * void IMatrixCollector::save_imatrix(int ncall) const { auto fname = m_params.out_file; - if (fname.empty()) { - fname = "imatrix.dat"; - } if (ncall > 0) { fname += ".at_"; @@ -498,7 +495,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) { const auto t_start = std::chrono::high_resolution_clock::now(); // clear the KV cache - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); llama_batch batch = llama_batch_init(n_batch, 0, 1); @@ -583,6 +580,8 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) { int main(int argc, char ** argv) { common_params params; + params.out_file = "imatrix.dat" ; + params.n_ctx = 512; params.logits_all = true; params.escape = false; diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp index 489a208b66..4e2f7b7270 100644 --- a/examples/infill/infill.cpp +++ b/examples/infill/infill.cpp @@ -332,8 +332,8 @@ int main(int argc, char ** argv) { LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n", n_past, n_left, n_ctx, params.n_keep, n_discard); - llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1 , params.n_keep + n_discard + 1); - llama_kv_cache_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard); + llama_kv_self_seq_rm (ctx, 0, params.n_keep + 1 , params.n_keep + n_discard + 1); + llama_kv_self_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard); n_past -= n_discard; diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index f518d02d38..cbcbfcee86 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -1578,7 +1578,7 @@ int main(int argc, char ** argv) { test t(inst, lmodel, ctx); - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); // cool off before the test if (params.delay) { @@ -1618,7 +1618,7 @@ int main(int argc, char ** argv) { } for (int i = 0; i < params.reps; i++) { - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); uint64_t t_start = get_time_ns(); diff --git a/examples/llama.android/llama/src/main/cpp/llama-android.cpp b/examples/llama.android/llama/src/main/cpp/llama-android.cpp index 2a73983a98..9654cd53cf 100644 --- a/examples/llama.android/llama/src/main/cpp/llama-android.cpp +++ b/examples/llama.android/llama/src/main/cpp/llama-android.cpp @@ -194,7 +194,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model( } batch->logits[batch->n_tokens - 1] = true; - llama_kv_cache_clear(context); + llama_kv_self_clear(context); const auto t_pp_start = ggml_time_us(); if (llama_decode(context, *batch) != 0) { @@ -206,7 +206,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model( LOGi("Benchmark text generation (tg)"); - llama_kv_cache_clear(context); + llama_kv_self_clear(context); const auto t_tg_start = ggml_time_us(); for (i = 0; i < tg; i++) { @@ -223,7 +223,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model( const auto t_tg_end = ggml_time_us(); - llama_kv_cache_clear(context); + llama_kv_self_clear(context); const auto t_pp = double(t_pp_end - t_pp_start) / 1000000.0; const auto t_tg = double(t_tg_end - t_tg_start) / 1000000.0; @@ -361,7 +361,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init( const auto tokens_list = common_tokenize(context, text, true, parse_special); auto n_ctx = llama_n_ctx(context); - auto n_kv_req = tokens_list.size() + (n_len - tokens_list.size()); + auto n_kv_req = tokens_list.size() + n_len; LOGi("n_len = %d, n_ctx = %d, n_kv_req = %d", n_len, n_ctx, n_kv_req); @@ -448,5 +448,5 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop( extern "C" JNIEXPORT void JNICALL Java_android_llama_cpp_LLamaAndroid_kv_1cache_1clear(JNIEnv *, jobject, jlong context) { - llama_kv_cache_clear(reinterpret_cast(context)); + llama_kv_self_clear(reinterpret_cast(context)); } diff --git a/examples/llama.swiftui/README.md b/examples/llama.swiftui/README.md index 5b0ee94720..bd7ce37747 100644 --- a/examples/llama.swiftui/README.md +++ b/examples/llama.swiftui/README.md @@ -16,7 +16,7 @@ Open `llama.swiftui.xcodeproj` project in Xcode and you should be able to build a simulator or a real device. To use the framework with a different project, the XCFramework can be added to the project by -adding `build-ios/llama.xcframework` by dragging and dropping it into the project navigator, or +adding `build-apple/llama.xcframework` by dragging and dropping it into the project navigator, or by manually selecting the framework in the "Frameworks, Libraries, and Embedded Content" section of the project settings. diff --git a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift index ee7141a663..f6e31abc93 100644 --- a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift +++ b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift @@ -210,7 +210,7 @@ actor LlamaContext { } batch.logits[Int(batch.n_tokens) - 1] = 1 // true - llama_kv_cache_clear(context) + llama_kv_self_clear(context) let t_pp_start = DispatchTime.now().uptimeNanoseconds / 1000; @@ -223,7 +223,7 @@ actor LlamaContext { // bench text generation - llama_kv_cache_clear(context) + llama_kv_self_clear(context) let t_tg_start = DispatchTime.now().uptimeNanoseconds / 1000; @@ -242,7 +242,7 @@ actor LlamaContext { let t_tg_end = DispatchTime.now().uptimeNanoseconds / 1000; - llama_kv_cache_clear(context) + llama_kv_self_clear(context) let t_pp = Double(t_pp_end - t_pp_start) / 1000000.0 let t_tg = Double(t_tg_end - t_tg_start) / 1000000.0 @@ -292,7 +292,7 @@ actor LlamaContext { func clear() { tokens_list.removeAll() temporary_invalid_cchars.removeAll() - llama_kv_cache_clear(context) + llama_kv_self_clear(context) } private func tokenize(text: String, add_bos: Bool) -> [llama_token] { diff --git a/examples/llava/CMakeLists.txt b/examples/llava/CMakeLists.txt index 319effd199..f275ce1ccd 100644 --- a/examples/llava/CMakeLists.txt +++ b/examples/llava/CMakeLists.txt @@ -51,6 +51,13 @@ install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_17) +set(TARGET llama-gemma3-cli) +add_executable(${TARGET} gemma3-cli.cpp) +set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-gemma3-cli) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_17) + set(TARGET llama-llava-clip-quantize-cli) add_executable(${TARGET} clip-quantize-cli.cpp) set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-llava-clip-quantize-cli) diff --git a/examples/llava/README-gemma3.md b/examples/llava/README-gemma3.md new file mode 100644 index 0000000000..20bf73fb5c --- /dev/null +++ b/examples/llava/README-gemma3.md @@ -0,0 +1,30 @@ +# Gemma 3 vision + +> [!IMPORTANT] +> +> This is very experimental, only used for demo purpose. + +## How to get mmproj.gguf? + +```bash +cd gemma-3-4b-it +python ../llama.cpp/examples/llava/gemma3_convert_encoder_to_gguf.py . + +# output file is mmproj.gguf +``` + +## How to run it? + +What you need: +- The text model GGUF, can be converted using `convert_hf_to_gguf.py` +- The mmproj file from step above +- An image file + +```bash +# build +cmake -B build +cmake --build build --target llama-gemma3-cli + +# run it +./build/bin/llama-gemma3-cli -m {text_model}.gguf --mmproj mmproj.gguf --image your_image.jpg +``` diff --git a/examples/llava/README-minicpmo2.6.md b/examples/llava/README-minicpmo2.6.md index 8f591506db..48c4232383 100644 --- a/examples/llava/README-minicpmo2.6.md +++ b/examples/llava/README-minicpmo2.6.md @@ -5,13 +5,25 @@ Currently, this readme only supports minicpm-omni's image capabilities, and we w Download [MiniCPM-o-2_6](https://huggingface.co/openbmb/MiniCPM-o-2_6) PyTorch model from huggingface to "MiniCPM-o-2_6" folder. + +### Build llama.cpp +Readme modification time: 20250206 + +If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md) + Clone llama.cpp: ```bash -git clone git@github.com:OpenBMB/llama.cpp.git +git clone https://github.com/ggerganov/llama.cpp cd llama.cpp -git checkout minicpm-omni ``` +Build llama.cpp using `CMake`: +```bash +cmake -B build +cmake --build build --config Release +``` + + ### Usage of MiniCPM-o 2.6 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-o-2_6-gguf) by us) @@ -22,25 +34,15 @@ python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM- python ./convert_hf_to_gguf.py ../MiniCPM-o-2_6/model # quantize int4 version -./llama-quantize ../MiniCPM-o-2_6/model/ggml-model-f16.gguf ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf Q4_K_M +./build/bin/llama-quantize ../MiniCPM-o-2_6/model/ggml-model-f16.gguf ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf Q4_K_M ``` -Build llama.cpp using `CMake`: -https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md - -```bash -cmake -B build -cmake --build build --config Release -``` Inference on Linux or Mac -``` +```bash # run f16 version -./llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" +./build/bin/llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" # run quantized int4 version -./llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" - -# or run in interactive mode -./llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i +./build/bin/llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" ``` diff --git a/examples/llava/README-minicpmv2.5.md b/examples/llava/README-minicpmv2.5.md index b0e72a0fa7..6bfe7abd16 100644 --- a/examples/llava/README-minicpmv2.5.md +++ b/examples/llava/README-minicpmv2.5.md @@ -4,13 +4,26 @@ Download [MiniCPM-Llama3-V-2_5](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5) PyTorch model from huggingface to "MiniCPM-Llama3-V-2_5" folder. + +### Build llama.cpp +Readme modification time: 20250206 + +If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md) + Clone llama.cpp: ```bash git clone https://github.com/ggml-org/llama.cpp cd llama.cpp ``` -### Usage +Build llama.cpp using `CMake`: +```bash +cmake -B build +cmake --build build --config Release +``` + + +### Usage of MiniCPM-Llama3-V 2.5 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5-gguf) by us) @@ -20,80 +33,15 @@ python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM- python ./convert_hf_to_gguf.py ../MiniCPM-Llama3-V-2_5/model # quantize int4 version -./llama-quantize ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf Q4_K_M +./build/bin/llama-quantize ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf Q4_K_M ``` -Build for Linux or Mac - -```bash -make -make llama-minicpmv-cli -``` Inference on Linux or Mac -``` +```bash # run f16 version -./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" +./build/bin/llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" # run quantized int4 version -./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" - -# or run in interactive mode -./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i -``` - -### Android - -#### Build on Android device using Termux -We found that build on Android device would bring better runtime performance, so we recommend to build on device. - -[Termux](https://github.com/termux/termux-app#installation) is a terminal app on Android device (no root required). - -Install tools in Termux: -``` -apt update && apt upgrade -y -apt install git make cmake -``` - -It's recommended to move your model inside the `~/` directory for best performance: -``` -cd storage/downloads -mv model.gguf ~/ -``` - -#### Building the Project using Android NDK -Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake. - -Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux: - -```bash -mkdir build-android -cd build-android -export NDK=/your_ndk_path -cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod .. -make -``` - -Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice). - -Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission: - -(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`) -``` -$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/ -$cd /data/data/com.termux/files/home/bin -$chmod +x ./* -``` - -Download models and push them to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/` - -``` -$mv /sdcard/llama.cpp/ggml-model-Q4_K_M.gguf /data/data/com.termux/files/home/model/ -$mv /sdcard/llama.cpp/mmproj-model-f16.gguf /data/data/com.termux/files/home/model/ -``` - -Now, you can start chatting: -``` -$cd /data/data/com.termux/files/home/bin -$./llama-minicpmv-cli -m ../model/ggml-model-Q4_K_M.gguf --mmproj ../model/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" +./build/bin/llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" ``` diff --git a/examples/llava/README-minicpmv2.6.md b/examples/llava/README-minicpmv2.6.md index c4be5e5dd6..2df39cdbac 100644 --- a/examples/llava/README-minicpmv2.6.md +++ b/examples/llava/README-minicpmv2.6.md @@ -4,13 +4,25 @@ Download [MiniCPM-V-2_6](https://huggingface.co/openbmb/MiniCPM-V-2_6) PyTorch model from huggingface to "MiniCPM-V-2_6" folder. + +### Build llama.cpp +Readme modification time: 20250206 + +If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md) + Clone llama.cpp: ```bash -git clone git@github.com:OpenBMB/llama.cpp.git +git clone https://github.com/ggerganov/llama.cpp cd llama.cpp -git checkout minicpmv-main ``` +Build llama.cpp using `CMake`: +```bash +cmake -B build +cmake --build build --config Release +``` + + ### Usage of MiniCPM-V 2.6 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) by us) @@ -21,87 +33,15 @@ python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM- python ./convert_hf_to_gguf.py ../MiniCPM-V-2_6/model # quantize int4 version -./llama-quantize ../MiniCPM-V-2_6/model/ggml-model-f16.gguf ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf Q4_K_M +./build/bin/llama-quantize ../MiniCPM-V-2_6/model/ggml-model-f16.gguf ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf Q4_K_M ``` -Build for Linux or Mac - -```bash -make -make llama-minicpmv-cli -``` Inference on Linux or Mac -``` +```bash # run f16 version -./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" +./build/bin/llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" # run quantized int4 version -./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" - -# or run in interactive mode -./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i -``` - -### Video -Install FFmpeg -``` -brew install ffmpeg -brew install pkg-config -``` - -### Android - -#### Build on Android device using Termux -We found that build on Android device would bring better runtime performance, so we recommend to build on device. - -[Termux](https://github.com/termux/termux-app#installation) is a terminal app on Android device (no root required). - -Install tools in Termux: -``` -apt update && apt upgrade -y -apt install git make cmake -``` - -It's recommended to move your model inside the `~/` directory for best performance: -``` -cd storage/downloads -mv model.gguf ~/ -``` - -#### Building the Project using Android NDK -Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake. - -Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux: - -```bash -mkdir build-android -cd build-android -export NDK=/your_ndk_path -cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod .. -make -``` - -Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice). - -Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission: - -(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`) -``` -$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/ -$cd /data/data/com.termux/files/home/bin -$chmod +x ./* -``` - -Download models and push them to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/` - -``` -$mv /sdcard/llama.cpp/ggml-model-Q4_K_M.gguf /data/data/com.termux/files/home/model/ -$mv /sdcard/llama.cpp/mmproj-model-f16.gguf /data/data/com.termux/files/home/model/ -``` - -Now, you can start chatting: -``` -$cd /data/data/com.termux/files/home/bin -$./llama-minicpmv-cli -m ../model/ggml-model-Q4_K_M.gguf --mmproj ../model/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" +./build/bin/llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" ``` diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index 76d4a78520..a1f050e39a 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -4,31 +4,12 @@ // Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch #include "clip.h" #include "ggml.h" +#include "ggml-cpp.h" #include "ggml-cpu.h" #include "ggml-alloc.h" #include "ggml-backend.h" #include "gguf.h" -//#ifdef GGML_USE_CUDA -//#include "ggml-cuda.h" -//#endif -// -//#ifdef GGML_USE_SYCL -//#include "ggml-sycl.h" -//#endif -// -//#ifdef GGML_USE_METAL -//#include "ggml-metal.h" -//#endif -// -//#ifdef GGML_USE_CANN -//#include "ggml-cann.h" -//#endif -// -//#ifdef GGML_USE_VULKAN -//#include "ggml-vulkan.h" -//#endif - #define STB_IMAGE_IMPLEMENTATION #include "stb_image.h" @@ -155,6 +136,8 @@ static std::string format(const char * fmt, ...) { #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s" #define TN_MVLM_PROJ_PEG "mm.model.peg.%d.%s" #define TN_IMAGE_NEWLINE "model.image_newline" +#define TN_MM_INP_PROJ "mm.input_projection.weight" // gemma3 +#define TN_MM_SOFT_EMB_N "mm.soft_emb_norm.weight" // gemma3 #define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k" #define TN_MINICPMV_QUERY "resampler.query" @@ -181,6 +164,7 @@ enum projector_type { PROJECTOR_TYPE_RESAMPLER, PROJECTOR_TYPE_GLM_EDGE, PROJECTOR_TYPE_MERGER, + PROJECTOR_TYPE_GEMMA3, PROJECTOR_TYPE_UNKNOWN, }; @@ -191,6 +175,7 @@ static std::map PROJECTOR_TYPE_NAMES = { { PROJECTOR_TYPE_RESAMPLER, "resampler"}, { PROJECTOR_TYPE_GLM_EDGE, "adapter"}, { PROJECTOR_TYPE_MERGER, "qwen2vl_merger"}, + { PROJECTOR_TYPE_GEMMA3, "gemma3"}, }; @@ -317,7 +302,7 @@ static projector_type clip_projector_type_from_string(const std::string & name) return kv.first; } } - return PROJECTOR_TYPE_UNKNOWN; + throw std::runtime_error(format("Unknown projector type: %s", name.c_str())); } #ifdef CLIP_DEBUG_FUNCTIONS @@ -574,6 +559,10 @@ struct clip_vision_model { struct ggml_tensor * mm_model_ln_kv_b; struct ggml_tensor * mm_model_ln_post_w; struct ggml_tensor * mm_model_ln_post_b; + + // gemma3 + struct ggml_tensor * mm_input_proj_w; + struct ggml_tensor * mm_soft_emb_norm_w; }; struct clip_ctx { @@ -588,7 +577,7 @@ struct clip_ctx { struct clip_vision_model vision_model; projector_type proj_type = PROJECTOR_TYPE_MLP; - int32_t max_feature_layer; + int32_t max_feature_layer; // unused in newer models like gemma3 float image_mean[3]; float image_std[3]; bool use_gelu = false; @@ -600,21 +589,209 @@ struct clip_ctx { bool has_post_norm = false; bool has_patch_bias = false; - struct gguf_context * ctx_gguf; - struct ggml_context * ctx_data; + struct gguf_context * ctx_gguf = nullptr; + struct ggml_context * ctx_data = nullptr; std::vector buf_compute_meta; - // memory buffers to evaluate the model - ggml_backend_buffer_t params_buffer = NULL; + std::vector backend_ptrs; + std::vector backend_buft; - ggml_backend_t backend = NULL; - ggml_gallocr_t compute_alloc = NULL; + ggml_backend_t backend = nullptr; + ggml_backend_t backend_cpu = nullptr; + ggml_backend_buffer_t buf = nullptr; - struct clip_image_size * load_image_size; + ggml_backend_sched_ptr sched; + + struct clip_image_size * load_image_size = nullptr; + + clip_ctx(clip_context_params & ctx_params) { + backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr); + backend = ctx_params.use_gpu + ? ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr) + : nullptr; + + if (backend) { + LOG_INF("%s: CLIP using %s backend\n", __func__, ggml_backend_name(backend)); + backend_ptrs.push_back(backend); + backend_buft.push_back(ggml_backend_get_default_buffer_type(backend)); + } else { + backend = backend_cpu; + LOG_INF("%s: CLIP using CPU backend\n", __func__); + } + + backend_ptrs.push_back(backend_cpu); + backend_buft.push_back(ggml_backend_get_default_buffer_type(backend_cpu)); + + sched.reset( + ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), 8192, false) + ); + } + + ~clip_ctx() { + ggml_free(ctx_data); + gguf_free(ctx_gguf); + ggml_backend_buffer_free(buf); + ggml_backend_free(backend); + if (backend_cpu != backend) { + ggml_backend_free(backend_cpu); + } + } }; -static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) { +static ggml_cgraph * clip_image_build_graph_siglip(clip_ctx * ctx, const clip_image_f32_batch * imgs) { + const auto & model = ctx->vision_model; + const auto & hparams = model.hparams; + + const int image_size = hparams.image_size; + int image_size_width = image_size; + int image_size_height = image_size; + + const int patch_size = hparams.patch_size; + const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size)); + const int hidden_size = hparams.hidden_size; + const int n_head = hparams.n_head; + const int d_head = hidden_size / n_head; + const int n_layer = hparams.n_layer; + const float eps = hparams.eps; + + GGML_ASSERT(imgs->size == 1); // batch_size == 1 + + struct ggml_init_params params = { + /*.mem_size =*/ ctx->buf_compute_meta.size(), + /*.mem_buffer =*/ ctx->buf_compute_meta.data(), + /*.no_alloc =*/ true, + }; + + struct ggml_context * ctx0 = ggml_init(params); + struct ggml_cgraph * gf = ggml_new_graph(ctx0); + + // input raw + struct ggml_tensor * inp_raw = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, 3); + ggml_set_name(inp_raw, "inp_raw"); + ggml_set_input(inp_raw); + + struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1); + inp = ggml_reshape_2d(ctx0, inp, num_patches, hidden_size); + inp = ggml_cont(ctx0, ggml_transpose(ctx0, inp)); + inp = ggml_add(ctx0, inp, model.patch_bias); + + // position embeddings + struct ggml_tensor * embeddings = ggml_add(ctx0, inp, model.position_embeddings); + + // loop over layers + for (int il = 0; il < n_layer; il++) { + struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states + + // layernorm1 + { + cur = ggml_norm(ctx0, cur, eps); + cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_1_w), model.layers[il].ln_1_b); + } + + // self-attention + { + + struct ggml_tensor * Q = + ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].q_w, cur), model.layers[il].q_b); + + Q = ggml_reshape_3d(ctx0, Q, d_head, n_head, num_patches); + Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3)); + + struct ggml_tensor * K = + ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].k_w, cur), model.layers[il].k_b); + + K = ggml_reshape_3d(ctx0, K, d_head, n_head, num_patches); + K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3)); + + struct ggml_tensor * V = + ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].v_w, cur), model.layers[il].v_b); + + V = ggml_reshape_3d(ctx0, V, d_head, n_head, num_patches); + V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3)); + + struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); + KQ = ggml_scale_inplace(ctx0, KQ, 1.0f / sqrtf((float)d_head)); + KQ = ggml_soft_max_inplace(ctx0, KQ); + + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ); + KQV = ggml_reshape_3d(ctx0, KQV, d_head, num_patches, n_head); + KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3); + + cur = ggml_cont_2d(ctx0, KQV, hidden_size, num_patches); + } + + // attention output + cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].o_w, cur), model.layers[il].o_b); + + // re-add the layer input, e.g., residual + cur = ggml_add(ctx0, cur, embeddings); + + embeddings = cur; // embeddings = residual, cur = hidden_states + + // layernorm2 + { + cur = ggml_norm(ctx0, cur, eps); + cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_2_w), model.layers[il].ln_2_b); + } + + cur = ggml_mul_mat(ctx0, model.layers[il].ff_i_w, cur); + cur = ggml_add(ctx0, cur, model.layers[il].ff_i_b); + + // siglip uses gelu + cur = ggml_gelu(ctx0, cur); + + cur = ggml_mul_mat(ctx0, model.layers[il].ff_o_w, cur); + cur = ggml_add(ctx0, cur, model.layers[il].ff_o_b); + + // residual 2 + cur = ggml_add(ctx0, embeddings, cur); + + embeddings = cur; + } + + // post-layernorm + if (ctx->has_post_norm) { + embeddings = ggml_norm(ctx0, embeddings, eps); + ggml_set_name(embeddings, "post_ln"); + + embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.post_ln_w), model.post_ln_b); + } + + if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) { + const int batch_size = 1; + const int mm_tokens_per_image = 256; // default value for gemma3 + const int tokens_per_side = sqrt(mm_tokens_per_image); + const int patches_per_image = sqrt(num_patches); + const int kernel_size = patches_per_image / tokens_per_side; + + embeddings = ggml_cont(ctx0, ggml_transpose(ctx0, embeddings)); + embeddings = ggml_reshape_4d(ctx0, embeddings, patches_per_image, patches_per_image, hidden_size, batch_size); + + // doing a pool2d to reduce the number of output tokens to 256 + embeddings = ggml_pool_2d(ctx0, embeddings, GGML_OP_POOL_AVG, kernel_size, kernel_size, kernel_size, kernel_size, 0, 0); + embeddings = ggml_reshape_3d(ctx0, embeddings, embeddings->ne[0] * embeddings->ne[0], hidden_size, batch_size); + embeddings = ggml_cont(ctx0, ggml_transpose(ctx0, embeddings)); + + // apply norm before projection + embeddings = ggml_rms_norm(ctx0, embeddings, eps); + embeddings = ggml_mul(ctx0, embeddings, model.mm_soft_emb_norm_w); + + // apply projection + embeddings = ggml_mul_mat(ctx0, + ggml_cont(ctx0, ggml_transpose(ctx0, model.mm_input_proj_w)), + embeddings); + } + + // build the graph + ggml_build_forward_expand(gf, embeddings); + + ggml_free(ctx0); + + return gf; +} + +static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) { if (!ctx->has_vision_encoder) { LOG_ERR("This gguf file seems to have no vision encoder\n"); return nullptr; @@ -1160,7 +1337,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 } else { GGML_ABORT("fatel error"); } - } else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) { + } + else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) { embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size * 4, num_positions / 4, batch_size); embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); @@ -1182,8 +1360,25 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 return gf; } +static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) { + if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) { + return clip_image_build_graph_siglip(ctx, imgs); + } else { + // TODO: we should have one build_* function per model + return clip_image_build_graph_legacy(ctx, imgs, load_image_size, is_inf); + } +} + // read and create ggml_context containing the tensors and their data struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { + return clip_init(fname, clip_context_params{ + /* use_gpu */ true, + /* verbosity */ verbosity, + }); +} + +struct clip_ctx * clip_init(const char * fname, struct clip_context_params ctx_params) { + int verbosity = ctx_params.verbosity; struct ggml_context * meta = NULL; struct gguf_init_params params = { @@ -1277,7 +1472,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { } } - clip_ctx * new_clip = new clip_ctx{}; + clip_ctx * new_clip = new clip_ctx(ctx_params); // update projector type { @@ -1296,36 +1491,6 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { } } -//#ifdef GGML_USE_CUDA -// new_clip->backend = ggml_backend_cuda_init(0); -// LOG_INF("%s: CLIP using CUDA backend\n", __func__); -//#endif -// -//#ifdef GGML_USE_METAL -// new_clip->backend = ggml_backend_metal_init(); -// LOG_INF("%s: CLIP using Metal backend\n", __func__); -//#endif -// -//#ifdef GGML_USE_CANN -// new_clip->backend = ggml_backend_cann_init(0); -// LOG_INF("%s: CLIP using CANN backend\n", __func__); -//#endif -// -//#ifdef GGML_USE_VULKAN -// new_clip->backend = ggml_backend_vk_init(0); -// LOG_INF("%s: CLIP using Vulkan backend\n", __func__); -//#endif -// -//#ifdef GGML_USE_SYCL -// new_clip->backend = ggml_backend_sycl_init(0); -// LOG_INF("%s: CLIP using SYCL backend\n", __func__); -//#endif - - if (!new_clip->backend) { - new_clip->backend = ggml_backend_cpu_init(); - LOG_INF("%s: CLIP using CPU backend\n", __func__); - } - // model size and capabilities { int idx = get_key_idx(ctx, KEY_HAS_TEXT_ENC); @@ -1363,8 +1528,12 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { GGML_ASSERT(new_clip->has_vision_encoder); GGML_ASSERT(!new_clip->has_text_encoder); - idx = get_key_idx(ctx, KEY_USE_GELU); - new_clip->use_gelu = gguf_get_val_bool(ctx, idx); + try { + idx = get_key_idx(ctx, KEY_USE_GELU); + new_clip->use_gelu = gguf_get_val_bool(ctx, idx); + } catch (std::runtime_error & /*e*/) { + new_clip->use_gelu = false; + } try { idx = get_key_idx(ctx, KEY_USE_SILU); @@ -1378,6 +1547,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { LOG_INF("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder); LOG_INF("%s: llava_projector: %d\n", __func__, new_clip->has_llava_projector); LOG_INF("%s: minicpmv_projector: %d\n", __func__, new_clip->has_minicpmv_projector); + LOG_INF("%s: minicpmv_version: %d\n", __func__, new_clip->minicpmv_version); LOG_INF("%s: glm_projector: %d\n", __func__, new_clip->has_glm_projector); LOG_INF("%s: model size: %.2f MB\n", __func__, model_size / 1024.0 / 1024.0); LOG_INF("%s: metadata size: %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0); @@ -1420,7 +1590,9 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { } // alloc memory and offload data - new_clip->params_buffer = ggml_backend_alloc_ctx_tensors(new_clip->ctx_data, new_clip->backend); + ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(new_clip->backend); + new_clip->buf = ggml_backend_alloc_ctx_tensors_from_buft(new_clip->ctx_data, buft); + ggml_backend_buffer_set_usage(new_clip->buf, GGML_BACKEND_BUFFER_USAGE_WEIGHTS); for (int i = 0; i < n_tensors; ++i) { const char * name = gguf_get_tensor_name(ctx, i); struct ggml_tensor * cur = ggml_get_tensor(new_clip->ctx_data, name); @@ -1433,7 +1605,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { return nullptr; } int num_bytes = ggml_nbytes(cur); - if (ggml_backend_buffer_is_host(new_clip->params_buffer)) { + if (ggml_backend_buft_is_host(buft)) { // for the CPU and Metal backend, we can read directly into the tensor fin.read(reinterpret_cast(cur->data), num_bytes); } else { @@ -1569,11 +1741,17 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { } try { - vision_model.patch_embeddings_0 = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD); + vision_model.patch_embeddings_0 = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD); + } catch(const std::exception& /*e*/) { + vision_model.patch_embeddings_0 = nullptr; + } + + try { vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v")); } catch(const std::exception& /*e*/) { - LOG_ERR("%s: failed to load vision model tensors\n", __func__); + vision_model.position_embeddings = nullptr; } + try { vision_model.patch_embeddings_1 = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD_1); } catch(const std::exception& /*e*/) { @@ -1684,6 +1862,10 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { vision_model.mm_1_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight")); vision_model.mm_1_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias")); } + else if (new_clip->proj_type == PROJECTOR_TYPE_GEMMA3) { + vision_model.mm_input_proj_w = get_tensor(new_clip->ctx_data, TN_MM_INP_PROJ); + vision_model.mm_soft_emb_norm_w = get_tensor(new_clip->ctx_data, TN_MM_SOFT_EMB_N); + } else { std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type]; throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); @@ -1719,14 +1901,21 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { // measure mem requirement and allocate { new_clip->buf_compute_meta.resize(GGML_DEFAULT_GRAPH_SIZE * ggml_tensor_overhead() + ggml_graph_overhead()); - new_clip->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_clip->backend)); clip_image_f32_batch batch; batch.size = 1; batch.data = nullptr; ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, nullptr, false); - ggml_gallocr_reserve(new_clip->compute_alloc, gf); - size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0); - LOG_INF("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0); + ggml_backend_sched_reserve(new_clip->sched.get(), gf); + for (size_t i = 0; i < new_clip->backend_ptrs.size(); ++i) { + ggml_backend_t backend = new_clip->backend_ptrs[i]; + ggml_backend_buffer_type_t buft = new_clip->backend_buft[i]; + size_t size = ggml_backend_sched_get_buffer_size(new_clip->sched.get(), backend); + if (size > 1) { + LOG_INF("%s: %10s compute buffer size = %8.2f MiB\n", __func__, + ggml_backend_buft_name(buft), + size / 1024.0 / 1024.0); + } + } } return new_clip; @@ -2218,7 +2407,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli return true; } - if (ctx->has_glm_projector) { + if (ctx->has_glm_projector || ctx->proj_type == PROJECTOR_TYPE_GEMMA3) { res_imgs->size = 1; res_imgs->data = new clip_image_f32[res_imgs->size]; clip_image_u8 resized_image; @@ -2407,12 +2596,6 @@ ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx) { } void clip_free(clip_ctx * ctx) { - ggml_free(ctx->ctx_data); - gguf_free(ctx->ctx_gguf); - - ggml_backend_buffer_free(ctx->params_buffer); - ggml_backend_free(ctx->backend); - ggml_gallocr_free(ctx->compute_alloc); delete ctx; } @@ -2608,8 +2791,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima } // build the inference graph + ggml_backend_sched_reset(ctx->sched.get()); ggml_cgraph * gf = clip_image_build_graph(ctx, imgs, ctx->load_image_size, true); - ggml_gallocr_alloc_graph(ctx->compute_alloc, gf); + ggml_backend_sched_alloc_graph(ctx->sched.get(), gf); // set inputs const auto & model = ctx->vision_model; @@ -2748,6 +2932,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions)); free(positions_data); } + else if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) { + // do nothing + } else { struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions"); @@ -2774,11 +2961,13 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima } } - if (ggml_backend_is_cpu(ctx->backend)) { - ggml_backend_cpu_set_n_threads(ctx->backend, n_threads); - } + ggml_backend_cpu_set_n_threads(ctx->backend_cpu, n_threads); - ggml_backend_graph_compute(ctx->backend, gf); + auto status = ggml_backend_sched_graph_compute(ctx->sched.get(), gf); + if (status != GGML_STATUS_SUCCESS) { + LOG_ERR("%s: ggml_backend_sched_graph_compute failed with error %d\n", __func__, status); + return false; + } // the last node is the embedding tensor struct ggml_tensor * embeddings = ggml_graph_node(gf, -1); @@ -2958,6 +3147,9 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) { if (ctx->proj_type == PROJECTOR_TYPE_MERGER) { return ctx->vision_model.mm_1_b->ne[0]; } + if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) { + return ctx->vision_model.mm_input_proj_w->ne[0]; + } std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type]; throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); diff --git a/examples/llava/clip.h b/examples/llava/clip.h index 002c419653..47059ca1b9 100644 --- a/examples/llava/clip.h +++ b/examples/llava/clip.h @@ -39,8 +39,15 @@ struct clip_image_f32_batch { size_t size; }; -CLIP_API struct clip_ctx * clip_model_load (const char * fname, int verbosity); -CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity); +struct clip_context_params { + bool use_gpu; + int verbosity; +}; + +// deprecated, use clip_init +CLIP_API struct clip_ctx * clip_model_load(const char * fname, int verbosity); + +CLIP_API struct clip_ctx * clip_init(const char * fname, struct clip_context_params ctx_params); CLIP_API void clip_free(struct clip_ctx * ctx); diff --git a/examples/llava/convert_image_encoder_to_gguf.py b/examples/llava/convert_image_encoder_to_gguf.py index de29687ec9..2949faec42 100644 --- a/examples/llava/convert_image_encoder_to_gguf.py +++ b/examples/llava/convert_image_encoder_to_gguf.py @@ -89,6 +89,7 @@ def bytes_to_unicode(): ap = argparse.ArgumentParser() ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True) ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16") +ap.add_argument('--bigendian', action="store_true", default=False, help="Model is executed on big-endian machine") ap.add_argument("--text-only", action="store_true", required=False, help="Save a text-only model. It can't be used to encode images") ap.add_argument("--vision-only", action="store_true", required=False, @@ -191,7 +192,7 @@ output_dir = args.output_dir if args.output_dir is not None else dir_model os.makedirs(output_dir, exist_ok=True) output_prefix = os.path.basename(output_dir).replace("ggml_", "") fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf") -fout = GGUFWriter(path=fname_out, arch="clip") +fout = GGUFWriter(path=fname_out, arch="clip", endianess=GGUFEndian.LITTLE if not args.bigendian else GGUFEndian.BIG) fout.add_bool("clip.has_text_encoder", has_text_encoder) fout.add_bool("clip.has_vision_encoder", has_vision_encoder) diff --git a/examples/llava/gemma3-cli.cpp b/examples/llava/gemma3-cli.cpp new file mode 100644 index 0000000000..c36bb2eda0 --- /dev/null +++ b/examples/llava/gemma3-cli.cpp @@ -0,0 +1,341 @@ +#include "arg.h" +#include "log.h" +#include "common.h" +#include "sampling.h" +#include "clip.h" +#include "stb_image.h" +#include "llama.h" +#include "ggml.h" +#include "console.h" + +#include +#include +#include + +#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) +#include +#include +#elif defined (_WIN32) +#define WIN32_LEAN_AND_MEAN +#ifndef NOMINMAX +#define NOMINMAX +#endif +#include +#include +#endif + +static bool g_is_generating = false; + +/** + * Please note that this is NOT a production-ready stuff. + * It is a playground for trying Gemma 3 vision capabilities. + * For contributors: please keep this code simple and easy to understand. + */ + +static void show_additional_info(int /*argc*/, char ** argv) { + LOG( + "Experimental CLI for using Gemma 3 vision model\n\n" + "Usage: %s [options] -m --mmproj --image -p \n\n" + " -m and --mmproj are required\n" + " --image and -p are optional, if NOT provided, the CLI will run in chat mode\n", + argv[0] + ); +} + +#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32) +static void sigint_handler(int signo) { + if (signo == SIGINT) { + if (g_is_generating) { + g_is_generating = false; + } else { + console::cleanup(); + LOG("\nInterrupted by user\n"); + _exit(130); + } + } +} +#endif + +struct gemma3_context { + struct clip_ctx * ctx_clip = NULL; + common_init_result llama_init; + + llama_model * model; + llama_context * lctx; + const llama_vocab * vocab; + llama_batch batch; + + int n_threads = 1; + llama_pos n_past = 0; + + gemma3_context(common_params & params) : llama_init(common_init_from_params(params)) { + model = llama_init.model.get(); + lctx = llama_init.context.get(); + vocab = llama_model_get_vocab(model); + n_threads = params.cpuparams.n_threads; + batch = llama_batch_init(params.n_batch, 0, 1); + init_clip_model(params); + } + + void init_clip_model(common_params & params) { + const char * clip_path = params.mmproj.c_str(); + ctx_clip = clip_model_load(clip_path, params.verbosity > 1); + } + + ~gemma3_context() { + clip_free(ctx_clip); + } +}; + +struct decode_embd_batch { + std::vector pos; + std::vector n_seq_id; + std::vector seq_id_0; + std::vector seq_ids; + std::vector logits; + llama_batch batch; + decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) { + pos .resize(n_tokens); + n_seq_id.resize(n_tokens); + seq_ids .resize(n_tokens + 1); + logits .resize(n_tokens); + seq_id_0.resize(1); + seq_id_0[0] = seq_id; + seq_ids [n_tokens] = nullptr; + batch = { + /*n_tokens =*/ n_tokens, + /*tokens =*/ nullptr, + /*embd =*/ embd, + /*pos =*/ pos.data(), + /*n_seq_id =*/ n_seq_id.data(), + /*seq_id =*/ seq_ids.data(), + /*logits =*/ logits.data(), + }; + for (int i = 0; i < n_tokens; i++) { + batch.pos [i] = pos_0 + i; + batch.n_seq_id[i] = 1; + batch.seq_id [i] = seq_id_0.data(); + batch.logits [i] = false; + } + } +}; + +static int eval_text(gemma3_context & ctx, std::string input, bool logits_last = false) { + llama_tokens tokens = common_tokenize(ctx.lctx, input, false, true); + common_batch_clear(ctx.batch); + for (llama_token & t : tokens) { + common_batch_add(ctx.batch, t, ctx.n_past++, {0}, false); + } + if (logits_last) { + ctx.batch.logits[ctx.batch.n_tokens - 1] = true; + } + // LOG("eval_text (n_tokens = %d): %s\n", (int)tokens.size(), input.c_str()); + if (llama_decode(ctx.lctx, ctx.batch)) { + LOG_ERR("Failed to decode text\n"); + return 1; + } + return 0; +} + +static int eval_image(gemma3_context & ctx, std::string & fname) { + std::vector image_embd_v; + int n_embd = llama_model_n_embd(ctx.model); + int n_tokens = 256; + image_embd_v.resize(n_tokens * n_embd); + + bool ok; + struct clip_image_u8 * img_u8 = clip_image_u8_init(); + ok = clip_image_load_from_file(fname.c_str(), img_u8); + if (!ok) { + LOG_ERR("Unable to load image %s\n", fname.c_str()); + clip_image_u8_free(img_u8); + return 2; // non-fatal error + } + + clip_image_f32_batch batch_f32; + ok = clip_image_preprocess(ctx.ctx_clip, img_u8, &batch_f32); + if (!ok) { + LOG_ERR("Unable to preprocess image\n"); + clip_image_f32_batch_free(&batch_f32); + clip_image_u8_free(img_u8); + return 1; + } + + int64_t t0 = ggml_time_ms(); + LOG("Encoding image %s\n", fname.c_str()); + ok = clip_image_batch_encode(ctx.ctx_clip, ctx.n_threads, &batch_f32, image_embd_v.data()); + if (!ok) { + LOG_ERR("Unable to encode image\n"); + clip_image_f32_batch_free(&batch_f32); + clip_image_u8_free(img_u8); + return 1; + } + LOG("Image encoded in %" PRId64 " ms\n", ggml_time_ms() - t0); + + clip_image_f32_batch_free(&batch_f32); + clip_image_u8_free(img_u8); + + // decode image embeddings + int64_t t1 = ggml_time_ms(); + eval_text(ctx, ""); + llama_set_causal_attn(ctx.lctx, false); + decode_embd_batch batch_img(image_embd_v.data(), n_tokens, ctx.n_past, 0); + if (llama_decode(ctx.lctx, batch_img.batch)) { + LOG_ERR("failed to decode image\n"); + return 1; + } + ctx.n_past += n_tokens; + llama_set_causal_attn(ctx.lctx, true); + eval_text(ctx, ""); + LOG("Image decoded in %" PRId64 " ms\n", ggml_time_ms() - t1); + return 0; +} + +static int generate_response(gemma3_context & ctx, common_sampler * smpl, int n_predict) { + for (int i = 0; i < n_predict; i++) { + if (i > n_predict || !g_is_generating) { + printf("\n"); + break; + } + + llama_token token_id = common_sampler_sample(smpl, ctx.lctx, -1); + common_sampler_accept(smpl, token_id, true); + + if (llama_vocab_is_eog(ctx.vocab, token_id)) { + printf("\n"); + break; // end of generation + } + + printf("%s", common_token_to_piece(ctx.lctx, token_id).c_str()); + fflush(stdout); + + // eval the token + common_batch_clear(ctx.batch); + common_batch_add(ctx.batch, token_id, ctx.n_past++, {0}, true); + if (llama_decode(ctx.lctx, ctx.batch)) { + LOG_ERR("failed to decode token\n"); + return 1; + } + } + return 0; +} + +int main(int argc, char ** argv) { + ggml_time_init(); + + common_params params; + params.sampling.temp = 0.2; // lower temp by default for better quality + + if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) { + return 1; + } + + common_init(); + + if (params.mmproj.empty()) { + show_additional_info(argc, argv); + return 1; + } + + gemma3_context ctx(params); + printf("%s: %s\n", __func__, params.model.c_str()); + + bool is_single_turn = !params.prompt.empty() && !params.image.empty(); + + struct common_sampler * smpl = common_sampler_init(ctx.model, params.sampling); + int n_predict = params.n_predict < 0 ? INT_MAX : params.n_predict; + + // ctrl+C handling + { +#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) + struct sigaction sigint_action; + sigint_action.sa_handler = sigint_handler; + sigemptyset (&sigint_action.sa_mask); + sigint_action.sa_flags = 0; + sigaction(SIGINT, &sigint_action, NULL); +#elif defined (_WIN32) + auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL { + return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false; + }; + SetConsoleCtrlHandler(reinterpret_cast(console_ctrl_handler), true); +#endif + } + + if (eval_text(ctx, "")) { + return 1; + } + + if (is_single_turn) { + g_is_generating = true; + if (eval_text(ctx, "user\n")) { + return 1; + } + for (auto & fname : params.image) { + if (eval_image(ctx, fname)) { + return 1; + } + } + if (eval_text(ctx, params.prompt + "model\n", true)) { + return 1; + } + if (generate_response(ctx, smpl, n_predict)) { + return 1; + } + + } else { + LOG("\n Running in chat mode, available commands:"); + LOG("\n /image load an image"); + LOG("\n /clear clear the chat history"); + LOG("\n /quit or /exit exit the program"); + LOG("\n"); + + if (eval_text(ctx, "user\n")) { + return 1; + } + + while (true) { + g_is_generating = false; + LOG("\n> "); + console::set_display(console::user_input); + std::string line; + console::readline(line, false); + console::set_display(console::reset); + line = string_strip(line); + if (line.empty()) { + continue; + } + if (line == "/quit" || line == "/exit") { + break; + } + if (line == "/clear") { + ctx.n_past = 0; + llama_kv_self_seq_rm(ctx.lctx, 0, 1, -1); // keep BOS + LOG("Chat history cleared\n\n"); + continue; + } + g_is_generating = true; + if (line.find("/image") == 0) { + std::string image = line.substr(7); + int res = eval_image(ctx, image); + if (res == 2) { + continue; // image not found + } + if (res) { + return 1; + } + continue; + } + if (eval_text(ctx, line + "model\n", true)) { + return 1; + } + if (generate_response(ctx, smpl, n_predict)) { + return 1; + } + if (eval_text(ctx, "user\n")) { + return 1; + } + } + } + + return 0; +} diff --git a/examples/llava/gemma3_convert_encoder_to_gguf.py b/examples/llava/gemma3_convert_encoder_to_gguf.py new file mode 100644 index 0000000000..241b526b9e --- /dev/null +++ b/examples/llava/gemma3_convert_encoder_to_gguf.py @@ -0,0 +1,307 @@ +import gguf +import argparse +import logging +import sys +import torch +import json +import os +import numpy as np +from typing import cast, ContextManager, Any, Iterator +from pathlib import Path +from torch import Tensor + +logger = logging.getLogger("gemma3-mmproj") + + +# (copied from convert_hf_to_gguf.py) +# tree of lazy tensors +class LazyTorchTensor(gguf.LazyBase): + _tensor_type = torch.Tensor + # to keep the type-checker happy + dtype: torch.dtype + shape: torch.Size + + # only used when converting a torch.Tensor to a np.ndarray + _dtype_map: dict[torch.dtype, type] = { + torch.float16: np.float16, + torch.float32: np.float32, + } + + # used for safetensors slices + # ref: https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/src/lib.rs#L1046 + # TODO: uncomment U64, U32, and U16, ref: https://github.com/pytorch/pytorch/issues/58734 + _dtype_str_map: dict[str, torch.dtype] = { + "F64": torch.float64, + "F32": torch.float32, + "BF16": torch.bfloat16, + "F16": torch.float16, + # "U64": torch.uint64, + "I64": torch.int64, + # "U32": torch.uint32, + "I32": torch.int32, + # "U16": torch.uint16, + "I16": torch.int16, + "U8": torch.uint8, + "I8": torch.int8, + "BOOL": torch.bool, + "F8_E4M3": torch.float8_e4m3fn, + "F8_E5M2": torch.float8_e5m2, + } + + def numpy(self) -> gguf.LazyNumpyTensor: + dtype = self._dtype_map[self.dtype] + return gguf.LazyNumpyTensor( + meta=gguf.LazyNumpyTensor.meta_with_dtype_and_shape(dtype, self.shape), + args=(self,), + func=(lambda s: s.numpy()) + ) + + @classmethod + def meta_with_dtype_and_shape(cls, dtype: torch.dtype, shape: tuple[int, ...]) -> Tensor: + return torch.empty(size=shape, dtype=dtype, device="meta") + + @classmethod + def from_safetensors_slice(cls, st_slice: Any) -> Tensor: + dtype = cls._dtype_str_map[st_slice.get_dtype()] + shape: tuple[int, ...] = tuple(st_slice.get_shape()) + lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(st_slice,), func=lambda s: s[:]) + return cast(torch.Tensor, lazy) + + @classmethod + def __torch_function__(cls, func, types, args=(), kwargs=None): + del types # unused + + if kwargs is None: + kwargs = {} + + if func is torch.Tensor.numpy: + return args[0].numpy() + + return cls._wrap_fn(func)(*args, **kwargs) + + +class Gemma3VisionTower: + hparams: dict + gguf_writer: gguf.GGUFWriter + fname_out: Path + ftype: gguf.LlamaFileType + + @staticmethod + def load_hparams(dir_model: Path): + with open(dir_model / "config.json", "r", encoding="utf-8") as f: + return json.load(f) + + @staticmethod + def get_model_part_names(dir_model: Path, prefix: str, suffix: str) -> list[str]: + part_names: list[str] = [] + for filename in os.listdir(dir_model): + if filename.startswith(prefix) and filename.endswith(suffix): + part_names.append(filename) + part_names.sort() + return part_names + + def __init__(self, + dir_model: Path, + fname_out: Path, + ftype: gguf.LlamaFileType, + is_big_endian: bool,): + hparams = Gemma3VisionTower.load_hparams(dir_model) + self.hparams = hparams + self.fname_out = fname_out + self.ftype = ftype + endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE + self.gguf_writer = gguf.GGUFWriter(path=None, arch="clip", endianess=endianess) + + text_config = hparams["text_config"] + vision_config = hparams["vision_config"] + + assert hparams["architectures"][0] == "Gemma3ForConditionalGeneration" + assert text_config is not None + assert vision_config is not None + + self.gguf_writer.add_string ("clip.projector_type", "gemma3") + self.gguf_writer.add_bool ("clip.has_text_encoder", False) + self.gguf_writer.add_bool ("clip.has_vision_encoder", True) + self.gguf_writer.add_bool ("clip.has_llava_projector", False) # legacy + self.gguf_writer.add_uint32 ("clip.vision.image_size", vision_config["image_size"]) + self.gguf_writer.add_uint32 ("clip.vision.patch_size", vision_config["patch_size"]) + self.gguf_writer.add_uint32 ("clip.vision.embedding_length", vision_config["hidden_size"]) + self.gguf_writer.add_uint32 ("clip.vision.feed_forward_length", vision_config["intermediate_size"]) + self.gguf_writer.add_uint32 ("clip.vision.projection_dim", text_config["hidden_size"]) + self.gguf_writer.add_uint32 ("clip.vision.block_count", vision_config["num_hidden_layers"]) + self.gguf_writer.add_uint32 ("clip.vision.attention.head_count", vision_config["num_attention_heads"]) + self.gguf_writer.add_float32("clip.vision.attention.layer_norm_epsilon", vision_config.get("layer_norm_eps", 1e-6)) + # default values taken from HF tranformers code + self.gguf_writer.add_array ("clip.vision.image_mean", [0.5, 0.5, 0.5]) + self.gguf_writer.add_array ("clip.vision.image_std", [0.5, 0.5, 0.5]) + self.gguf_writer.add_bool ("clip.use_gelu", True) + + # load tensors + for name, data_torch in self.get_tensors(dir_model): + # convert any unsupported data types to float32 + if data_torch.dtype not in (torch.float16, torch.float32): + data_torch = data_torch.to(torch.float32) + self.add_tensor(name, data_torch) + + def get_tensors(self, dir_model: Path) -> Iterator[tuple[str, Tensor]]: + part_names = Gemma3VisionTower.get_model_part_names(dir_model, "model", ".safetensors") + tensor_names_from_parts: set[str] = set() + for part_name in part_names: + logger.info(f"gguf: loading model part '{part_name}'") + from safetensors import safe_open + ctx = cast(ContextManager[Any], safe_open(dir_model / part_name, framework="pt", device="cpu")) + with ctx as model_part: + tensor_names_from_parts.update(model_part.keys()) + + for name in model_part.keys(): + data = model_part.get_slice(name) + data = LazyTorchTensor.from_safetensors_slice(data) + yield name, data + + def add_tensor(self, name: str, data_torch: Tensor): + is_1d = len(data_torch.shape) == 1 + is_embd = ".embeddings." in name + old_dtype = data_torch.dtype + can_quantize = not is_1d and not is_embd + data_qtype = gguf.GGMLQuantizationType.F32 + + # this is to support old checkpoint + # TODO: remove this when we have the final model + name = name.replace("vision_model.vision_model.", "vision_tower.vision_model.") + name = name.replace("multimodal_projector.", "multi_modal_projector.") + + # filter only vision tensors + if not name.startswith("vision_tower.vision_model.") and not name.startswith("multi_modal_projector."): + return + # prefix + name = name.replace("vision_tower.vision_model.encoder.layers.", "v.blk.") + name = name.replace("vision_tower.vision_model.", "v.") + # projector and input embd + name = name.replace(".embeddings.patch_embedding.", ".patch_embd.") + name = name.replace(".embeddings.position_embedding.", ".position_embd.") + name = name.replace( + "multi_modal_projector.mm_input_projection_weight", + "mm.input_projection.weight" + ) + name = name.replace( + "multi_modal_projector.mm_soft_emb_norm.weight", + "mm.soft_emb_norm.weight" + ) + name = name.replace("post_layernorm.", "post_ln.") + # each block + name = name.replace(".self_attn.k_proj.", ".attn_k.") + name = name.replace(".self_attn.v_proj.", ".attn_v.") + name = name.replace(".self_attn.q_proj.", ".attn_q.") + name = name.replace(".self_attn.out_proj.", ".attn_out.") + name = name.replace(".layer_norm1.", ".ln1.") + name = name.replace(".layer_norm2.", ".ln2.") + name = name.replace(".mlp.fc1.", ".ffn_down.") + name = name.replace(".mlp.fc2.", ".ffn_up.") + + if can_quantize: + if self.ftype == gguf.LlamaFileType.ALL_F32: + data_qtype = gguf.GGMLQuantizationType.F32 + elif self.ftype == gguf.LlamaFileType.MOSTLY_F16: + data_qtype = gguf.GGMLQuantizationType.F16 + elif self.ftype == gguf.LlamaFileType.MOSTLY_BF16: + data_qtype = gguf.GGMLQuantizationType.BF16 + elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0: + data_qtype = gguf.GGMLQuantizationType.Q8_0 + else: + raise ValueError(f"Unsupported file type: {self.ftype}") + + # corrent norm value ; only this "soft_emb_norm" need to be corrected as it's part of Gemma projector + # the other norm values are part of SigLIP model, and they are already correct + # ref code: Gemma3RMSNorm + if "soft_emb_norm.weight" in name: + logger.info(f"Correcting norm value for '{name}'") + data_torch = data_torch + 1 + + data = data_torch.numpy() + + try: + data = gguf.quants.quantize(data, data_qtype) + except Exception as e: + logger.error(f"Error quantizing tensor '{name}': {e}, fallback to F16") + data_qtype = gguf.GGMLQuantizationType.F16 + data = gguf.quants.quantize(data, data_qtype) + + # reverse shape to make it similar to the internal ggml dimension order + shape_str = f"{{{', '.join(str(n) for n in reversed(data_torch.shape))}}}" + logger.info(f"{f'%-32s' % f'{name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}") + + self.gguf_writer.add_tensor(name, data, raw_dtype=data_qtype) + + def write(self): + self.gguf_writer.write_header_to_file(path=self.fname_out) + self.gguf_writer.write_kv_data_to_file() + self.gguf_writer.write_tensors_to_file(progress=True) + self.gguf_writer.close() + +def parse_args() -> argparse.Namespace: + parser = argparse.ArgumentParser( + description="Convert Gemma 3 vision tower safetensors to GGUF format",) + parser.add_argument( + "--outfile", type=Path, default="mmproj.gguf", + help="path to write to", + ) + parser.add_argument( + "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0"], default="f16", + help="output format", + ) + parser.add_argument( + "--bigendian", action="store_true", + help="model is executed on big endian machine", + ) + parser.add_argument( + "model", type=Path, + help="directory containing model file", + nargs="?", + ) + parser.add_argument( + "--verbose", action="store_true", + help="increase output verbosity", + ) + + args = parser.parse_args() + if args.model is None: + parser.error("the following arguments are required: model") + return args + + +def main() -> None: + args = parse_args() + + if args.verbose: + logging.basicConfig(level=logging.DEBUG) + else: + logging.basicConfig(level=logging.INFO) + + dir_model = args.model + + if not dir_model.is_dir(): + logger.error(f'Error: {args.model} is not a directory') + sys.exit(1) + + ftype_map: dict[str, gguf.LlamaFileType] = { + "f32": gguf.LlamaFileType.ALL_F32, + "f16": gguf.LlamaFileType.MOSTLY_F16, + "bf16": gguf.LlamaFileType.MOSTLY_BF16, + "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0, + } + + logger.info(f"Loading model: {dir_model.name}") + + with torch.inference_mode(): + gemma3_vision_tower = Gemma3VisionTower( + dir_model=dir_model, + fname_out=args.outfile, + ftype=ftype_map[args.outtype], + is_big_endian=args.bigendian, + ) + gemma3_vision_tower.write() + + +if __name__ == '__main__': + main() + diff --git a/examples/llava/minicpmv-cli.cpp b/examples/llava/minicpmv-cli.cpp index 53d902d616..12f536cf5c 100644 --- a/examples/llava/minicpmv-cli.cpp +++ b/examples/llava/minicpmv-cli.cpp @@ -86,7 +86,11 @@ static struct clip_ctx * clip_init_context(common_params * params) { if (prompt.empty()) { prompt = "describe the image in detail."; } - auto * ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1); + struct clip_context_params clip_params = { + /* use_gpu */ params->n_gpu_layers != 0, + /* verbosity */ params->verbosity, + }; + auto * ctx_clip = clip_init(clip_path, clip_params); return ctx_clip; } @@ -148,19 +152,34 @@ static void process_image(struct llava_context * ctx_llava, struct llava_image_e process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++); eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); if (num_image_embeds > 1) { - size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip); - eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); - for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) { - for (size_t j = 0; j < num_image_embeds_col; ++j) { - eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); - process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++); - eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); - if (j == num_image_embeds_col - 1) { - eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false); + if (has_minicpmv_projector == 2) { + size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip); + eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); + for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) { + for (size_t j = 0; j < num_image_embeds_col; ++j) { + eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); + process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++); + eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); + if (j == num_image_embeds_col - 1) { + eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false); + } + } + } + eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); + } + else if (has_minicpmv_projector == 3 || has_minicpmv_projector == 4) { + size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip); + for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) { + for (size_t j = 0; j < num_image_embeds_col; ++j) { + eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); + process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++); + eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); + if (j == num_image_embeds_col - 1) { + eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false); + } } } } - eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); } LOG_INF("%s: image token past: %d\n", __func__, n_past); } diff --git a/examples/llava/minicpmv-convert-image-encoder-to-gguf.py b/examples/llava/minicpmv-convert-image-encoder-to-gguf.py index 9b196757f0..cfe0961f98 100644 --- a/examples/llava/minicpmv-convert-image-encoder-to-gguf.py +++ b/examples/llava/minicpmv-convert-image-encoder-to-gguf.py @@ -597,7 +597,6 @@ elif args.minicpmv_projector is not None: fname_middle = "mmproj-" has_text_encoder = False has_minicpmv_projector = True - minicpmv_version = 4 elif args.vision_only: fname_middle = "vision-" has_text_encoder = False diff --git a/examples/lookahead/lookahead.cpp b/examples/lookahead/lookahead.cpp index b9e8de694c..7df20aee17 100644 --- a/examples/lookahead/lookahead.cpp +++ b/examples/lookahead/lookahead.cpp @@ -96,7 +96,7 @@ int main(int argc, char ** argv) { llama_decode(ctx, llama_batch_get_one(&inp.back(), 1)); for (int s = 1; s < W + G + 1; ++s) { - llama_kv_cache_seq_cp(ctx, 0, s, -1, -1); + llama_kv_self_seq_cp(ctx, 0, s, -1, -1); } const auto t_enc_end = ggml_time_us(); @@ -438,17 +438,17 @@ int main(int argc, char ** argv) { // KV cache management // if no verification token matched, we simply remove all cells from this batch -> no fragmentation - llama_kv_cache_seq_rm(ctx, -1, n_past, -1); + llama_kv_self_seq_rm(ctx, -1, n_past, -1); if (seq_id_best != 0) { // if a verification token matched, we keep the best sequence and remove the rest // this leads to some KV cache fragmentation - llama_kv_cache_seq_keep(ctx, seq_id_best); - llama_kv_cache_seq_cp (ctx, seq_id_best, 0, -1, -1); - llama_kv_cache_seq_rm (ctx, seq_id_best, -1, -1); + llama_kv_self_seq_keep(ctx, seq_id_best); + llama_kv_self_seq_cp (ctx, seq_id_best, 0, -1, -1); + llama_kv_self_seq_rm (ctx, seq_id_best, -1, -1); for (int s = 1; s < W + G + 1; ++s) { - llama_kv_cache_seq_cp(ctx, 0, s, -1, -1); + llama_kv_self_seq_cp(ctx, 0, s, -1, -1); } } } diff --git a/examples/lookup/lookup.cpp b/examples/lookup/lookup.cpp index dbd0444ec8..4ae93b2a5e 100644 --- a/examples/lookup/lookup.cpp +++ b/examples/lookup/lookup.cpp @@ -192,7 +192,7 @@ int main(int argc, char ** argv){ // KV cache management // clean the cache of draft tokens that weren't accepted - llama_kv_cache_seq_rm(ctx, 0, n_past, -1); + llama_kv_self_seq_rm(ctx, 0, n_past, -1); common_batch_clear(batch_tgt); common_batch_add(batch_tgt, draft[0], n_past, { 0 }, true); diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 4e0c69473b..fd7410a646 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -354,7 +354,7 @@ int main(int argc, char ** argv) { } // remove any "future" tokens that we might have inherited from the previous session - llama_kv_cache_seq_rm(ctx, -1, n_matching_session_tokens, -1); + llama_kv_self_seq_rm(ctx, -1, n_matching_session_tokens, -1); } LOG_DBG("recalculate the cached logits (check): embd_inp.size() %zu, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu\n", @@ -602,8 +602,8 @@ int main(int argc, char ** argv) { LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n", n_past, n_left, n_ctx, params.n_keep, n_discard); - llama_kv_cache_seq_rm (ctx, 0, params.n_keep , params.n_keep + n_discard); - llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard); + llama_kv_self_seq_rm (ctx, 0, params.n_keep , params.n_keep + n_discard); + llama_kv_self_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard); n_past -= n_discard; @@ -626,9 +626,9 @@ int main(int argc, char ** argv) { LOG_DBG("div: [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n); LOG_DBG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd); - llama_kv_cache_seq_add(ctx, 0, ga_i, n_past, ib*bd); - llama_kv_cache_seq_div(ctx, 0, ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n); - llama_kv_cache_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd, dd); + llama_kv_self_seq_add(ctx, 0, ga_i, n_past, ib*bd); + llama_kv_self_seq_div(ctx, 0, ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n); + llama_kv_self_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd, dd); n_past -= bd; diff --git a/examples/parallel/parallel.cpp b/examples/parallel/parallel.cpp index be18909ed4..588632f043 100644 --- a/examples/parallel/parallel.cpp +++ b/examples/parallel/parallel.cpp @@ -202,7 +202,7 @@ int main(int argc, char ** argv) { // assign the system KV cache to all parallel sequences for (int32_t i = 1; i <= n_clients; ++i) { - llama_kv_cache_seq_cp(ctx, 0, i, -1, -1); + llama_kv_self_seq_cp(ctx, 0, i, -1, -1); } LOG_INF("\n"); @@ -234,9 +234,9 @@ int main(int argc, char ** argv) { if (batch.n_tokens == 0) { // all sequences have ended - clear the entire KV cache for (int i = 1; i <= n_clients; ++i) { - llama_kv_cache_seq_rm(ctx, i, -1, -1); + llama_kv_self_seq_rm(ctx, i, -1, -1); // but keep the system prompt - llama_kv_cache_seq_cp(ctx, 0, i, -1, -1); + llama_kv_self_seq_cp(ctx, 0, i, -1, -1); } LOG_INF("%s: clearing the KV cache\n", __func__); @@ -372,8 +372,8 @@ int main(int argc, char ** argv) { } // delete only the generated part of the sequence, i.e. keep the system prompt in the cache - llama_kv_cache_seq_rm(ctx, client.id + 1, -1, -1); - llama_kv_cache_seq_cp(ctx, 0, client.id + 1, -1, -1); + llama_kv_self_seq_rm(ctx, client.id + 1, -1, -1); + llama_kv_self_seq_cp(ctx, 0, client.id + 1, -1, -1); const auto t_main_end = ggml_time_us(); diff --git a/examples/passkey/passkey.cpp b/examples/passkey/passkey.cpp index fa85190518..ea3a6c1fca 100644 --- a/examples/passkey/passkey.cpp +++ b/examples/passkey/passkey.cpp @@ -133,11 +133,11 @@ int main(int argc, char ** argv) { const int ib = i/n_batch - 1; const int bd = n_batch_grp*(n_grp - 1); - llama_kv_cache_seq_add (ctx, 0, n_past - n_batch, n_past, ib*bd); - llama_kv_cache_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp); - llama_kv_cache_update (ctx); + llama_kv_self_seq_add (ctx, 0, n_past - n_batch, n_past, ib*bd); + llama_kv_self_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp); + llama_kv_self_update (ctx); - n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1; + n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1; } common_batch_clear(batch); @@ -167,12 +167,12 @@ int main(int argc, char ** argv) { LOG_INF("%s: shifting KV cache with %d\n", __func__, n_discard); - llama_kv_cache_seq_rm (ctx, 0, n_keep , n_keep + n_discard); - llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard); - //llama_kv_cache_defrag (ctx); - llama_kv_cache_update (ctx); + llama_kv_self_seq_rm (ctx, 0, n_keep , n_keep + n_discard); + llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard); + //llama_kv_self_defrag (ctx); + llama_kv_self_update (ctx); - n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1; + n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1; common_batch_clear(batch); @@ -198,12 +198,12 @@ int main(int argc, char ** argv) { if (n_discard > 0) { LOG_INF("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard); - llama_kv_cache_seq_rm (ctx, 0, n_keep , n_keep + n_discard); - llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard); - //llama_kv_cache_defrag (ctx); - llama_kv_cache_update (ctx); + llama_kv_self_seq_rm (ctx, 0, n_keep , n_keep + n_discard); + llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard); + //llama_kv_self_defrag (ctx); + llama_kv_self_update (ctx); - n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1; + n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1; } } diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 5d07421e82..8c413f7d66 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -361,7 +361,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params const auto t_start = std::chrono::high_resolution_clock::now(); // clear the KV cache - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); llama_batch batch = llama_batch_init(n_batch, 0, 1); @@ -547,7 +547,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params & const auto t_start = std::chrono::high_resolution_clock::now(); // clear the KV cache - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); for (int j = 0; j < num_batches; ++j) { const int batch_start = start + j * n_batch; @@ -924,7 +924,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) { return; } - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); // decode all tasks [i0, i1) if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) { @@ -1203,7 +1203,7 @@ static void winogrande_score(llama_context * ctx, const common_params & params) return; } - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); // decode all tasks [i0, i1) if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) { @@ -1575,7 +1575,7 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par return; } - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); // decode all tasks [i0, i1) if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) { @@ -1765,7 +1765,7 @@ static void kl_divergence(llama_context * ctx, const common_params & params) { } // clear the KV cache - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); llama_batch batch = llama_batch_init(n_batch, 0, 1); diff --git a/examples/quantize-stats/quantize-stats.cpp b/examples/quantize-stats/quantize-stats.cpp index bd2f734670..dd07ab9b37 100644 --- a/examples/quantize-stats/quantize-stats.cpp +++ b/examples/quantize-stats/quantize-stats.cpp @@ -1,6 +1,6 @@ #include "ggml.h" #include "llama.h" -#include "llama-context.h" +#include "llama-model.h" #include "common.h" #include @@ -328,7 +328,7 @@ int main(int argc, char ** argv) { } } - const auto & tensors = llama_internal_get_tensor_map(ctx); + const auto & tensors = llama_internal_get_tensor_map(model); // check layer tensors int included_layers = 0; diff --git a/examples/retrieval/retrieval.cpp b/examples/retrieval/retrieval.cpp index 2439022a22..0efe20d4b3 100644 --- a/examples/retrieval/retrieval.cpp +++ b/examples/retrieval/retrieval.cpp @@ -83,7 +83,7 @@ static void batch_add_seq(llama_batch & batch, const std::vector & toke static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) { // clear previous kv_cache values (irrelevant for embeddings) - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); // run model LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq); diff --git a/examples/run/run.cpp b/examples/run/run.cpp index 38407d5190..437f2533e5 100644 --- a/examples/run/run.cpp +++ b/examples/run/run.cpp @@ -891,7 +891,7 @@ static int apply_chat_template(const struct common_chat_templates * tmpls, Llama // Function to tokenize the prompt static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt, std::vector & prompt_tokens, const LlamaData & llama_data) { - const bool is_first = llama_get_kv_cache_used_cells(llama_data.context.get()) == 0; + const bool is_first = llama_kv_self_used_cells(llama_data.context.get()) == 0; const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true); prompt_tokens.resize(n_prompt_tokens); @@ -907,7 +907,7 @@ static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt // Check if we have enough space in the context to evaluate this batch static int check_context_size(const llama_context_ptr & ctx, const llama_batch & batch) { const int n_ctx = llama_n_ctx(ctx.get()); - const int n_ctx_used = llama_get_kv_cache_used_cells(ctx.get()); + const int n_ctx_used = llama_kv_self_used_cells(ctx.get()); if (n_ctx_used + batch.n_tokens > n_ctx) { printf(LOG_COL_DEFAULT "\n"); printe("context size exceeded\n"); diff --git a/examples/save-load-state/save-load-state.cpp b/examples/save-load-state/save-load-state.cpp index cf7cbd8159..760ebbbf08 100644 --- a/examples/save-load-state/save-load-state.cpp +++ b/examples/save-load-state/save-load-state.cpp @@ -15,7 +15,7 @@ int main(int argc, char ** argv) { return 1; } - print_build_info(); + common_init(); if (params.n_predict < 0) { params.n_predict = 16; @@ -196,7 +196,7 @@ int main(int argc, char ** argv) { fprintf(stderr, "%s : seq 0 copied, %zd bytes\n", __func__, ncopy); // erase whole kv - llama_kv_cache_clear(ctx3); + llama_kv_self_clear(ctx3); fprintf(stderr, "%s : kv cache cleared\n", __func__); // restore kv into seq 1 diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 2a526b0e7a..71e053b202 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -384,8 +384,9 @@ struct server_task { SRV_DBG("Grammar trigger token: %d (`%s`)\n", token, word.c_str()); common_grammar_trigger trigger; trigger.type = COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN; - trigger.value = (llama_token) token; - params.sampling.grammar_triggers.push_back(trigger); + trigger.value = word; + trigger.token = token; + params.sampling.grammar_triggers.push_back(std::move(trigger)); } else { SRV_DBG("Grammar trigger word: `%s`\n", word.c_str()); params.sampling.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, word}); @@ -750,7 +751,10 @@ struct server_task_result_cmpl_final : server_task_result { {"name", tc.name}, {"arguments", tc.arguments}, }}, - {"id", tc.id}, + // Some templates generate and require an id (sometimes in a very specific format, e.g. Mistral Nemo). + // We only generate a random id for the ones that don't generate one by themselves + // (they also won't get to see it as their template likely doesn't use it, so it's all for the client) + {"id", tc.id.empty() ? gen_tool_call_id() : tc.id}, }); } message["tool_calls"] = tool_calls; @@ -1312,7 +1316,7 @@ struct server_slot { return task_type == SERVER_TASK_TYPE_EMBEDDING || task_type == SERVER_TASK_TYPE_RERANK; } - bool can_batch_with(server_slot & other_slot) { + bool can_batch_with(server_slot & other_slot) const { return is_non_causal() == other_slot.is_non_causal() && are_lora_equal(lora, other_slot.lora); } @@ -1900,6 +1904,7 @@ struct server_context { try { common_chat_format_example(chat_templates.get(), params.use_jinja); } catch (const std::exception & e) { + SRV_WRN("%s: Chat template parsing error: %s\n", __func__, e.what()); SRV_WRN("%s: The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses\n", __func__); chat_templates = common_chat_templates_init(model, "chatml"); } @@ -2035,6 +2040,18 @@ struct server_context { return ret; } + bool can_be_detokenized(const struct llama_context * ctx, const std::vector & tokens) { + const llama_model * model = llama_get_model(ctx); + const llama_vocab * vocab = llama_model_get_vocab(model); + const int32_t n_vocab = llama_vocab_n_tokens(vocab); + for (const auto & token : tokens) { + if (token < 0 || token >= n_vocab) { + return false; + } + } + return true; + } + bool launch_slot_with_task(server_slot & slot, const server_task & task) { slot.reset(); slot.id_task = task.id; @@ -2049,6 +2066,11 @@ struct server_context { slot.lora = task.params.lora; } + bool can_detokenize = can_be_detokenized(ctx, slot.prompt_tokens); + if (!can_detokenize) { + send_error(task, "Prompt contains invalid tokens", ERROR_TYPE_INVALID_REQUEST); + return false; + } SLT_DBG(slot, "launching slot : %s\n", safe_json_to_str(slot.to_json()).c_str()); if (slot.n_predict > 0 && slot.params.n_predict > slot.n_predict) { @@ -2091,7 +2113,7 @@ struct server_context { SRV_DBG("%s", "clearing KV cache\n"); // clear the entire KV cache - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); clean_kv_cache = false; } @@ -2156,14 +2178,6 @@ struct server_context { } if (slot.has_new_line) { - // if we have already seen a new line, we stop after a certain time limit - if (slot.params.t_max_predict_ms > 0 && (ggml_time_us() - slot.t_start_generation > 1000.0f*slot.params.t_max_predict_ms)) { - slot.stop = STOP_TYPE_LIMIT; - slot.has_next_token = false; - - SLT_DBG(slot, "stopped by time limit, n_decoded = %d, t_max_predict_ms = %d ms\n", slot.n_decoded, (int) slot.params.t_max_predict_ms); - } - // require that each new line has a whitespace prefix (i.e. indentation) of at least slot.params.n_indent if (slot.params.n_indent > 0) { // check the current indentation @@ -2202,6 +2216,14 @@ struct server_context { // check if there is a new line in the generated text if (result.text_to_send.find('\n') != std::string::npos) { slot.has_new_line = true; + + // if we have seen a new line, we stop after a certain time limit, but only upon another new line + if (slot.params.t_max_predict_ms > 0 && (ggml_time_us() - slot.t_start_generation > 1000.0f*slot.params.t_max_predict_ms)) { + slot.stop = STOP_TYPE_LIMIT; + slot.has_next_token = false; + + SLT_DBG(slot, "stopped by time limit, n_decoded = %d, t_max_predict_ms = %d ms\n", slot.n_decoded, (int) slot.params.t_max_predict_ms); + } } // if context shift is disabled, we stop when it reaches the context limit @@ -2633,8 +2655,8 @@ struct server_context { res->n_tasks_deferred = queue_tasks.queue_tasks_deferred.size(); res->t_start = metrics.t_start; - res->kv_cache_tokens_count = llama_get_kv_cache_token_count(ctx); - res->kv_cache_used_cells = llama_get_kv_cache_used_cells(ctx); + res->kv_cache_tokens_count = llama_kv_self_n_tokens(ctx); + res->kv_cache_used_cells = llama_kv_self_used_cells(ctx); res->n_prompt_tokens_processed_total = metrics.n_prompt_tokens_processed_total; res->t_prompt_processing_total = metrics.t_prompt_processing_total; @@ -2750,7 +2772,7 @@ struct server_context { // Erase token cache const size_t n_erased = slot->cache_tokens.size(); - llama_kv_cache_seq_rm(ctx, slot->id, -1, -1); + llama_kv_self_seq_rm(ctx, slot->id, -1, -1); slot->cache_tokens.clear(); auto res = std::make_unique(); @@ -2818,8 +2840,8 @@ struct server_context { SLT_WRN(slot, "slot context shift, n_keep = %d, n_left = %d, n_discard = %d\n", n_keep, n_left, n_discard); - llama_kv_cache_seq_rm (ctx, slot.id, n_keep , n_keep + n_discard); - llama_kv_cache_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past, -n_discard); + llama_kv_self_seq_rm (ctx, slot.id, n_keep , n_keep + n_discard); + llama_kv_self_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past, -n_discard); if (slot.params.cache_prompt) { for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++) { @@ -3010,8 +3032,8 @@ struct server_context { const int64_t kv_shift = (int64_t) head_p - (int64_t) head_c; - llama_kv_cache_seq_rm (ctx, slot.id, head_p, head_c); - llama_kv_cache_seq_add(ctx, slot.id, head_c, head_c + n_match, kv_shift); + llama_kv_self_seq_rm (ctx, slot.id, head_p, head_c); + llama_kv_self_seq_add(ctx, slot.id, head_c, head_c + n_match, kv_shift); for (size_t i = 0; i < n_match; i++) { slot.cache_tokens[head_p + i] = slot.cache_tokens[head_c + i]; @@ -3049,9 +3071,9 @@ struct server_context { } // keep only the common part - if (!llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1)) { + if (!llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1)) { // could not partially delete (likely using a non-Transformer model) - llama_kv_cache_seq_rm(ctx, slot.id, -1, -1); + llama_kv_self_seq_rm(ctx, slot.id, -1, -1); // there is no common part left slot.n_past = 0; @@ -3291,7 +3313,7 @@ struct server_context { slot.cache_tokens.push_back(id); slot.cache_tokens.insert(slot.cache_tokens.end(), ids.begin(), ids.end() - 1); - llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1); + llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1); for (size_t i = 0; i < ids.size(); ++i) { completion_token_output result; diff --git a/examples/server/tests/unit/test_tool_call.py b/examples/server/tests/unit/test_tool_call.py index 25bddbaee7..569c2a1f8e 100755 --- a/examples/server/tests/unit/test_tool_call.py +++ b/examples/server/tests/unit/test_tool_call.py @@ -92,6 +92,7 @@ def do_test_completion_with_required_tool_tiny(server: ServerProcess, tool: dict assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}' tool_call = tool_calls[0] assert choice["message"].get("content") in (None, ""), f'Expected no content in {choice["message"]}' + assert len(tool_call.get("id", "")) > 0, f'Expected non empty tool call id in {tool_call}' expected_function_name = "python" if tool["type"] == "code_interpreter" else tool["function"]["name"] assert expected_function_name == tool_call["function"]["name"] actual_arguments = tool_call["function"]["arguments"] @@ -373,6 +374,7 @@ def do_test_weather(server: ServerProcess, **kwargs): tool_call = tool_calls[0] # assert choice["message"].get("content") in (None, ""), f'Expected no content in {choice["message"]}' assert tool_call["function"]["name"] == WEATHER_TOOL["function"]["name"], f'Expected weather tool call, got {tool_call["function"]["name"]}' + assert len(tool_call.get("id", "")) > 0, f'Expected non empty tool call id in {tool_call}' actual_arguments = json.loads(tool_call["function"]["arguments"]) assert 'location' in actual_arguments, f"location not found in {json.dumps(actual_arguments)}" location = actual_arguments["location"] @@ -596,6 +598,7 @@ def do_test_hello_world(server: ServerProcess, **kwargs): tool_call = tool_calls[0] # assert choice["message"].get("content") in (None, ""), f'Expected no content in {choice["message"]}' assert tool_call["function"]["name"] == PYTHON_TOOL["function"]["name"] + assert len(tool_call.get("id", "")) > 0, f'Expected non empty tool call id in {tool_call}' actual_arguments = json.loads(tool_call["function"]["arguments"]) assert 'code' in actual_arguments, f"code not found in {json.dumps(actual_arguments)}" code = actual_arguments["code"] diff --git a/examples/server/tests/utils.py b/examples/server/tests/utils.py index ec2d8ec558..30aa866095 100644 --- a/examples/server/tests/utils.py +++ b/examples/server/tests/utils.py @@ -302,7 +302,7 @@ class ServerPreset: server.model_hf_repo = "ggml-org/models" server.model_hf_file = "tinyllamas/stories260K.gguf" server.model_alias = "tinyllama-2" - server.n_ctx = 256 + server.n_ctx = 512 server.n_batch = 32 server.n_slots = 2 server.n_predict = 64 diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp index 393e3927c7..36ad276fd3 100644 --- a/examples/server/utils.hpp +++ b/examples/server/utils.hpp @@ -435,6 +435,10 @@ static std::string gen_chatcmplid() { return "chatcmpl-" + random_string(); } +static std::string gen_tool_call_id() { + return random_string(); +} + // // other common utils // diff --git a/examples/simple-chat/simple-chat.cpp b/examples/simple-chat/simple-chat.cpp index c5534cc13e..84f4159737 100644 --- a/examples/simple-chat/simple-chat.cpp +++ b/examples/simple-chat/simple-chat.cpp @@ -98,7 +98,7 @@ int main(int argc, char ** argv) { auto generate = [&](const std::string & prompt) { std::string response; - const bool is_first = llama_get_kv_cache_used_cells(ctx) == 0; + const bool is_first = llama_kv_self_used_cells(ctx) == 0; // tokenize the prompt const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true); @@ -113,7 +113,7 @@ int main(int argc, char ** argv) { while (true) { // check if we have enough space in the context to evaluate this batch int n_ctx = llama_n_ctx(ctx); - int n_ctx_used = llama_get_kv_cache_used_cells(ctx); + int n_ctx_used = llama_kv_self_used_cells(ctx); if (n_ctx_used + batch.n_tokens > n_ctx) { printf("\033[0m\n"); fprintf(stderr, "context size exceeded\n"); diff --git a/examples/speculative-simple/speculative-simple.cpp b/examples/speculative-simple/speculative-simple.cpp index 403ba2dd21..a5d2bc9d09 100644 --- a/examples/speculative-simple/speculative-simple.cpp +++ b/examples/speculative-simple/speculative-simple.cpp @@ -217,7 +217,7 @@ int main(int argc, char ** argv) { { LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past); - llama_kv_cache_seq_rm(ctx_tgt, 0, n_past, -1); + llama_kv_self_seq_rm(ctx_tgt, 0, n_past, -1); } if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) { diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp index c7ccea50db..bfddc67e03 100644 --- a/examples/speculative/speculative.cpp +++ b/examples/speculative/speculative.cpp @@ -420,14 +420,14 @@ int main(int argc, char ** argv) { { LOG_DBG("keeping sequence %d, n_past_tgt = %d, n_past_dft = %d\n", s_keep, n_past_tgt, n_past_dft); - llama_kv_cache_seq_keep(ctx_dft, s_keep); - llama_kv_cache_seq_cp (ctx_dft, s_keep, 0, -1, -1); - llama_kv_cache_seq_keep(ctx_dft, 0); + llama_kv_self_seq_keep(ctx_dft, s_keep); + llama_kv_self_seq_cp (ctx_dft, s_keep, 0, -1, -1); + llama_kv_self_seq_keep(ctx_dft, 0); - llama_kv_cache_seq_rm (ctx_tgt, s_keep, n_past_tgt, -1); - llama_kv_cache_seq_keep(ctx_tgt, s_keep); - llama_kv_cache_seq_cp (ctx_tgt, s_keep, 0, -1, -1); - llama_kv_cache_seq_keep(ctx_tgt, 0); + llama_kv_self_seq_rm (ctx_tgt, s_keep, n_past_tgt, -1); + llama_kv_self_seq_keep(ctx_tgt, s_keep); + llama_kv_self_seq_cp (ctx_tgt, s_keep, 0, -1, -1); + llama_kv_self_seq_keep(ctx_tgt, 0); } for (int s = 0; s < n_seq_dft; ++s) { @@ -444,7 +444,7 @@ int main(int argc, char ** argv) { common_batch_clear(batch_dft); common_batch_add (batch_dft, token_id, n_past_dft, { 0 }, true); - llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1); + llama_kv_self_seq_rm(ctx_dft, 0, n_past_dft, -1); // LOG_DBG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str()); llama_decode(ctx_dft, batch_dft); @@ -503,8 +503,8 @@ int main(int argc, char ** argv) { if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_draft_split) { LOG_DBG("splitting seq %3d into %3d\n", s, n_seq_cur); - llama_kv_cache_seq_rm(ctx_dft, n_seq_cur, -1, -1); - llama_kv_cache_seq_cp(ctx_dft, s, n_seq_cur, -1, -1); + llama_kv_self_seq_rm(ctx_dft, n_seq_cur, -1, -1); + llama_kv_self_seq_cp(ctx_dft, s, n_seq_cur, -1, -1); // all previous tokens from this branch are now also part of the new branch for (int t = 0; t < batch_tgt.n_tokens; ++t) { @@ -585,9 +585,9 @@ int main(int argc, char ** argv) { // evaluate the target model on the drafted tokens { - llama_kv_cache_seq_keep(ctx_tgt, 0); + llama_kv_self_seq_keep(ctx_tgt, 0); for (int s = 1; s < n_seq_dft; ++s) { - llama_kv_cache_seq_cp(ctx_tgt, 0, s, -1, -1); + llama_kv_self_seq_cp(ctx_tgt, 0, s, -1, -1); } // LOG_DBG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str()); diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt index 5c314999db..f24a8692cb 100644 --- a/ggml/CMakeLists.txt +++ b/ggml/CMakeLists.txt @@ -106,6 +106,7 @@ option(GGML_CPU_KLEIDIAI "ggml: use KleidiAI optimized kernels if applicable option(GGML_AVX "ggml: enable AVX" ${INS_ENB}) option(GGML_AVX_VNNI "ggml: enable AVX-VNNI" OFF) option(GGML_AVX2 "ggml: enable AVX2" ${INS_ENB}) +option(GGML_BMI2 "ggml: enable BMI2" ${INS_ENB}) option(GGML_AVX512 "ggml: enable AVX512F" OFF) option(GGML_AVX512_VBMI "ggml: enable AVX512-VBMI" OFF) option(GGML_AVX512_VNNI "ggml: enable AVX512-VNNI" OFF) @@ -195,6 +196,8 @@ option(GGML_OPENCL "ggml: use OpenCL" option(GGML_OPENCL_PROFILING "ggml: use OpenCL profiling (increases overhead)" OFF) option(GGML_OPENCL_EMBED_KERNELS "ggml: embed kernels" ON) option(GGML_OPENCL_USE_ADRENO_KERNELS "ggml: use optimized kernels for Adreno" ON) +set (GGML_OPENCL_TARGET_VERSION "300" CACHE STRING + "gmml: OpenCL API version to target") # toolchain for vulkan-shaders-gen set (GGML_VULKAN_SHADERS_GEN_TOOLCHAIN "" CACHE FILEPATH "ggml: toolchain file for vulkan-shaders-gen") diff --git a/ggml/include/ggml-cpu.h b/ggml/include/ggml-cpu.h index b48cc560e5..f5e11f1e10 100644 --- a/ggml/include/ggml-cpu.h +++ b/ggml/include/ggml-cpu.h @@ -80,6 +80,7 @@ extern "C" { GGML_BACKEND_API int ggml_cpu_has_avx (void); GGML_BACKEND_API int ggml_cpu_has_avx_vnni (void); GGML_BACKEND_API int ggml_cpu_has_avx2 (void); + GGML_BACKEND_API int ggml_cpu_has_bmi2 (void); GGML_BACKEND_API int ggml_cpu_has_f16c (void); GGML_BACKEND_API int ggml_cpu_has_fma (void); GGML_BACKEND_API int ggml_cpu_has_avx512 (void); diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index 9cf5a3885b..e4b8c52e35 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -236,7 +236,7 @@ add_library(ggml target_link_libraries(ggml PUBLIC ggml-base) if (CMAKE_SYSTEM_NAME MATCHES "Linux") - target_link_libraries(ggml PRIVATE dl) + target_link_libraries(ggml PRIVATE dl stdc++fs) endif() function(ggml_add_backend_library backend) @@ -289,7 +289,7 @@ function(ggml_add_cpu_backend_variant tag_name) set(GGML_CPU_TAG_NAME ${tag_name}) # other: OPENMP LLAMAFILE CPU_HBM foreach (feat NATIVE - AVX AVX2 AVX_VNNI FMA F16C + AVX AVX2 BMI2 AVX_VNNI FMA F16C AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8 AMX_BF16) set(GGML_${feat} OFF) @@ -309,13 +309,13 @@ if (GGML_CPU_ALL_VARIANTS) message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL") endif() ggml_add_cpu_backend_variant(sandybridge AVX) - ggml_add_cpu_backend_variant(haswell AVX F16C AVX2 FMA) - ggml_add_cpu_backend_variant(skylakex AVX F16C AVX2 FMA AVX512) - ggml_add_cpu_backend_variant(icelake AVX F16C AVX2 FMA AVX512 AVX512_VBMI AVX512_VNNI) - ggml_add_cpu_backend_variant(alderlake AVX F16C AVX2 FMA AVX_VNNI) + ggml_add_cpu_backend_variant(haswell AVX F16C AVX2 BMI2 FMA) + ggml_add_cpu_backend_variant(skylakex AVX F16C AVX2 BMI2 FMA AVX512) + ggml_add_cpu_backend_variant(icelake AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI) + ggml_add_cpu_backend_variant(alderlake AVX F16C AVX2 BMI2 FMA AVX_VNNI) if (NOT MSVC) # MSVC doesn't support AMX - ggml_add_cpu_backend_variant(sapphirerapids AVX F16C AVX2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8) + ggml_add_cpu_backend_variant(sapphirerapids AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8) endif() elseif (GGML_CPU) ggml_add_cpu_backend_variant_impl("") diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp index 3c5b2d9852..9c6cc32786 100644 --- a/ggml/src/ggml-backend-reg.cpp +++ b/ggml/src/ggml-backend-reg.cpp @@ -80,7 +80,14 @@ namespace fs = std::filesystem; static std::string path_str(const fs::path & path) { std::string u8path; try { +#if defined(__cpp_lib_char8_t) + // C++20 and later: u8string() returns std::u8string + std::u8string u8str = path.u8string(); + u8path = std::string(reinterpret_cast(u8str.c_str())); +#else + // C++17: u8string() returns std::string u8path = path.u8string(); +#endif } catch (...) { } return u8path; @@ -497,7 +504,7 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent, search_paths.push_back(get_executable_path()); search_paths.push_back(fs::current_path()); } else { - search_paths.push_back(user_search_path); + search_paths.push_back(fs::u8path(user_search_path)); } int best_score = 0; @@ -511,9 +518,9 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent, fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied); for (const auto & entry : dir_it) { if (entry.is_regular_file()) { - auto filename = entry.path().filename().native(); - auto ext = entry.path().extension().native(); - if (filename.find(file_prefix) == 0 && ext == file_extension) { + auto filename = entry.path().filename(); + auto ext = entry.path().extension(); + if (filename.native().find(file_prefix) == 0 && ext == file_extension) { dl_handle_ptr handle { dl_load_library(entry) }; if (!handle && !silent) { GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path_str(entry.path()).c_str()); @@ -544,7 +551,7 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent, // try to load the base backend for (const auto & search_path : search_paths) { fs::path filename = backend_filename_prefix().native() + name_path.native() + backend_filename_extension().native(); - fs::path path = search_path.native() + filename.native(); + fs::path path = search_path / filename; if (fs::exists(path)) { return get_reg().load_backend(path, silent); } diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt index f8836ed61b..d6c4a9c299 100644 --- a/ggml/src/ggml-cpu/CMakeLists.txt +++ b/ggml/src/ggml-cpu/CMakeLists.txt @@ -219,6 +219,10 @@ function(ggml_add_cpu_backend_variant_impl tag_name) if (GGML_AVX_VNNI) list(APPEND ARCH_DEFINITIONS __AVXVNNI__ GGML_AVX_VNNI) endif() + if (GGML_BMI2) + # MSVC does not define macro __BMI2__ + list(APPEND ARCH_DEFINITIONS __BMI2__ GGML_BMI2) + endif() else () if (GGML_NATIVE) list(APPEND ARCH_FLAGS -march=native) @@ -233,6 +237,10 @@ function(ggml_add_cpu_backend_variant_impl tag_name) list(APPEND ARCH_FLAGS -mfma) list(APPEND ARCH_DEFINITIONS GGML_FMA) endif() + if (GGML_BMI2) + list(APPEND ARCH_FLAGS -mbmi2) + list(APPEND ARCH_DEFINITIONS GGML_BMI2) + endif() if (GGML_AVX) list(APPEND ARCH_FLAGS -mavx) list(APPEND ARCH_DEFINITIONS GGML_AVX) diff --git a/ggml/src/ggml-cpu/cpu-feats-x86.cpp b/ggml/src/ggml-cpu/cpu-feats-x86.cpp index e8133d411f..902ee43466 100644 --- a/ggml/src/ggml-cpu/cpu-feats-x86.cpp +++ b/ggml/src/ggml-cpu/cpu-feats-x86.cpp @@ -278,6 +278,10 @@ static int ggml_backend_cpu_x86_score() { if (!is.SSE42()) { return 0; } score += 1<<2; #endif +#ifdef GGML_BMI2 + if (!is.BMI2()) { return 0; } + score += 1<<3; +#endif #ifdef GGML_AVX if (!is.AVX()) { return 0; } score += 1<<4; diff --git a/ggml/src/ggml-cpu/ggml-cpu-quants.c b/ggml/src/ggml-cpu/ggml-cpu-quants.c index c30ac03185..8c7dbd1ccb 100644 --- a/ggml/src/ggml-cpu/ggml-cpu-quants.c +++ b/ggml/src/ggml-cpu/ggml-cpu-quants.c @@ -11362,10 +11362,19 @@ void ggml_vec_dot_iq1_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const __m256i sumi = _mm256_setzero_si256(); int sumi1 = 0; for (int ib = 0; ib < QK_K/32; ib += 2) { +#ifdef __BMI2__ + const uint64_t packed_idx1 = _pdep_u64(*(const uint32_t *)qs, 0x00ff00ff00ff00ffULL) | _pdep_u64(qh[ib], 0x700070007000700ULL); + const uint64_t packed_idx2 = _pdep_u64(*(const uint32_t *)(qs + 4), 0x00ff00ff00ff00ffULL) | _pdep_u64(qh[ib + 1], 0x700070007000700ULL); + const uint16_t *idx1 = (const uint16_t *)(&packed_idx1); + const uint16_t *idx2 = (const uint16_t *)(&packed_idx2); + const __m256i q1b_1 = _mm256_set_epi64x(iq1s_grid[idx1[3]], iq1s_grid[idx1[2]], iq1s_grid[idx1[1]], iq1s_grid[idx1[0]]); + const __m256i q1b_2 = _mm256_set_epi64x(iq1s_grid[idx2[3]], iq1s_grid[idx2[2]], iq1s_grid[idx2[1]], iq1s_grid[idx2[0]]); +#else const __m256i q1b_1 = _mm256_set_epi64x(iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)], iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)]); const __m256i q1b_2 = _mm256_set_epi64x(iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)], iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)]); +#endif qs += 8; const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; @@ -11711,6 +11720,10 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const const __m256i mask = _mm256_set1_epi16(0x7); const __m256i mone = _mm256_set1_epi16(1); + const __m256i mone8 = _mm256_set1_epi8(1); + const __m256i mtwo8 = _mm256_set1_epi8(2); + // VPSHUFB cannot cross 128-bit lanes so odd shifts go to upper half. + const __m256i scales_shift = _mm256_set_epi64x(9, 3, 6, 0); __m256 accum1 = _mm256_setzero_ps(); __m256 accum2 = _mm256_setzero_ps(); @@ -11722,10 +11735,33 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const const uint16_t * sc = (const uint16_t *)x[i].scales; scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); + // Extract 3-bit scales (16 values) + __m256i scales = _mm256_set1_epi64x(*(const uint64_t*)sc); + scales = _mm256_srlv_epi64(scales, scales_shift); + scales = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scales, mask), 1), mone); + + // Indices to repeat each scale 8 times. + __m256i scales_idx1 = _mm256_set1_epi16(0x0100); + __m256i scales_idx2 = _mm256_add_epi8(scales_idx1, _mm256_set1_epi8(8)); __m256i sumi1 = _mm256_setzero_si256(); __m256i sumi2 = _mm256_setzero_si256(); for (int ib = 0; ib < QK_K/32; ib += 2) { +#ifdef __BMI2__ + const uint64_t packed_idx1 = _pdep_u64(*(const uint32_t *)qs, 0x00ff00ff00ff00ffULL) + | _pdep_u64(*(const uint16_t*)(qh) & 0x7777, 0xf000f000f000f00ULL); + const uint64_t packed_idx2 = _pdep_u64(*(const uint32_t *)(qs + 4), 0x00ff00ff00ff00ffULL) + | _pdep_u64(*(const uint16_t*)(qh + 2) & 0x7777, 0xf000f000f000f00ULL); + const uint16_t *idx1 = (const uint16_t *)(&packed_idx1); + const uint16_t *idx2 = (const uint16_t *)(&packed_idx2); + const __m256i q1b_1 = _mm256_set_epi64x(iq1s_grid[idx1[3]], iq1s_grid[idx1[2]], iq1s_grid[idx1[1]], iq1s_grid[idx1[0]]); + const __m256i q1b_2 = _mm256_set_epi64x(iq1s_grid[idx2[3]], iq1s_grid[idx2[2]], iq1s_grid[idx2[1]], iq1s_grid[idx2[0]]); + + // Convert signs to bytes 0x81 (negative) or 0x01 (positive) + const uint64_t delta_sign = _pdep_u64(*(const uint32_t*)(qh) & 0x88888888, 0xf0f0f0f0f0f0f0f0ULL); + const __m256i delta1 = _mm256_or_si256(mone8, _mm256_cvtepi8_epi64(_mm_set1_epi32(delta_sign))); + const __m256i delta2 = _mm256_or_si256(mone8, _mm256_cvtepi8_epi64(_mm_set1_epi32(delta_sign >> 32))); +#else const __m256i q1b_1 = _mm256_set_epi64x( iq1s_grid[qs[3] | (((uint16_t)qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | (((uint16_t)qh[1] << 8) & 0x700)], iq1s_grid[qs[1] | (((uint16_t)qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | (((uint16_t)qh[0] << 8) & 0x700)] @@ -11734,11 +11770,6 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const iq1s_grid[qs[7] | (((uint16_t)qh[3] << 4) & 0x700)], iq1s_grid[qs[6] | (((uint16_t)qh[3] << 8) & 0x700)], iq1s_grid[qs[5] | (((uint16_t)qh[2] << 4) & 0x700)], iq1s_grid[qs[4] | (((uint16_t)qh[2] << 8) & 0x700)] ); - const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; - const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; - - const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1); - const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2); const __m256i delta1 = _mm256_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101, qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101, @@ -11748,15 +11779,21 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const qh[3] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101, qh[2] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101, qh[2] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101); +#endif + const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; + const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; - const __m256i dot3 = mul_add_epi8(delta1, q8b_1); - const __m256i dot4 = mul_add_epi8(delta2, q8b_2); + const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1); + const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2); + const __m256i dot3 = _mm256_maddubs_epi16(mone8, _mm256_sign_epi8(q8b_1, delta1)); + const __m256i dot4 = _mm256_maddubs_epi16(mone8, _mm256_sign_epi8(q8b_2, delta2)); - __m256i scale1 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 3), _mm_set1_epi16(sc[ib/2] >> 0)); - __m256i scale2 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 9), _mm_set1_epi16(sc[ib/2] >> 6)); + __m256i scale1 = _mm256_shuffle_epi8(scales, scales_idx1); + __m256i scale2 = _mm256_shuffle_epi8(scales, scales_idx2); + + scales_idx1 = _mm256_add_epi8(scales_idx1, mtwo8); + scales_idx2 = _mm256_add_epi8(scales_idx2, mtwo8); - scale1 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale1, mask), 1), mone); - scale2 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale2, mask), 1), mone); const __m256i p1 = _mm256_madd_epi16(dot1, scale1); const __m256i p2 = _mm256_madd_epi16(dot2, scale2); const __m256i p3 = _mm256_madd_epi16(dot3, scale1); diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c index 2a5463fdff..f2ab4c5d69 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.c +++ b/ggml/src/ggml-cpu/ggml-cpu.c @@ -6648,6 +6648,135 @@ static void ggml_compute_forward_repeat_back( // ggml_compute_forward_concat +static void ggml_compute_forward_concat_any( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + const size_t len = ggml_type_size(src0->type); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int32_t dim = ggml_get_op_params_i32(dst, 0); + + GGML_ASSERT(dim >= 0 && dim < 4); + + int64_t o[4] = {0, 0, 0, 0}; + o[dim] = src0->ne[dim]; + + const char * x; + + // TODO: smarter multi-theading + for (int i3 = 0; i3 < ne3; i3++) { + for (int i2 = ith; i2 < ne2; i2 += nth) { + for (int i1 = 0; i1 < ne1; i1++) { + for (int i0 = 0; i0 < ne0; i0++) { + if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { + x = (const char *)src0->data + (i0 )*nb00 + (i1 )*nb01 + (i2 )*nb02 + (i3 )*nb03; + } else { + x = (const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13; + } + + char * y = (char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3; + + memcpy(y, x, len); + } + } + } + } +} + +static void ggml_compute_forward_concat_i8( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_type_size(src0->type) == sizeof(int8_t)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int32_t dim = ggml_get_op_params_i32(dst, 0); + + GGML_ASSERT(dim >= 0 && dim < 4); + + int64_t o[4] = {0, 0, 0, 0}; + o[dim] = src0->ne[dim]; + + const int8_t * x; + + // TODO: smarter multi-theading + for (int i3 = 0; i3 < ne3; i3++) { + for (int i2 = ith; i2 < ne2; i2 += nth) { + for (int i1 = 0; i1 < ne1; i1++) { + for (int i0 = 0; i0 < ne0; i0++) { + if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { + x = (const int8_t *) ((const char *)src0->data + (i0 )*nb00 + (i1 )*nb01 + (i2 )*nb02 + (i3 )*nb03); + } else { + x = (const int8_t *) ((const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13); + } + + int8_t * y = (int8_t *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3); + + *y = *x; + } + } + } + } +} + +static void ggml_compute_forward_concat_f16( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_type_size(src0->type) == sizeof(ggml_fp16_t)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int32_t dim = ggml_get_op_params_i32(dst, 0); + + GGML_ASSERT(dim >= 0 && dim < 4); + + int64_t o[4] = {0, 0, 0, 0}; + o[dim] = src0->ne[dim]; + + const ggml_fp16_t * x; + + // TODO: smarter multi-theading + for (int i3 = 0; i3 < ne3; i3++) { + for (int i2 = ith; i2 < ne2; i2 += nth) { + for (int i1 = 0; i1 < ne1; i1++) { + for (int i0 = 0; i0 < ne0; i0++) { + if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { + x = (const ggml_fp16_t *) ((const char *)src0->data + (i0 )*nb00 + (i1 )*nb01 + (i2 )*nb02 + (i3 )*nb03); + } else { + x = (const ggml_fp16_t *) ((const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13); + } + + ggml_fp16_t * y = (ggml_fp16_t *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3); + + *y = *x; + } + } + } + } +} + static void ggml_compute_forward_concat_f32( const struct ggml_compute_params * params, struct ggml_tensor * dst) { @@ -6655,7 +6784,7 @@ static void ggml_compute_forward_concat_f32( const struct ggml_tensor * src0 = dst->src[0]; const struct ggml_tensor * src1 = dst->src[1]; - GGML_ASSERT(src0->nb[0] == sizeof(float)); + GGML_ASSERT(ggml_type_size(src0->type) == sizeof(float)); const int ith = params->ith; const int nth = params->nth; @@ -6698,6 +6827,16 @@ static void ggml_compute_forward_concat( const struct ggml_tensor * src0 = dst->src[0]; switch (src0->type) { + case GGML_TYPE_F16: + case GGML_TYPE_BF16: + case GGML_TYPE_I16: + { + ggml_compute_forward_concat_f16(params, dst); + } break; + case GGML_TYPE_I8: + { + ggml_compute_forward_concat_i8(params, dst); + } break; case GGML_TYPE_F32: case GGML_TYPE_I32: { @@ -6705,7 +6844,7 @@ static void ggml_compute_forward_concat( } break; default: { - GGML_ABORT("fatal error"); + ggml_compute_forward_concat_any(params, dst); } } } @@ -15440,6 +15579,14 @@ int ggml_cpu_has_amx_int8(void) { #endif } +int ggml_cpu_has_bmi2(void) { +#if defined(__BMI2__) + return 1; +#else + return 0; +#endif +} + int ggml_cpu_has_fma(void) { #if defined(__FMA__) return 1; diff --git a/ggml/src/ggml-cpu/ggml-cpu.cpp b/ggml/src/ggml-cpu/ggml-cpu.cpp index a84203f29f..09f8382b98 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.cpp +++ b/ggml/src/ggml-cpu/ggml-cpu.cpp @@ -511,6 +511,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r if (ggml_cpu_has_fma()) { features.push_back({ "FMA", "1" }); } + if (ggml_cpu_has_bmi2()) { + features.push_back({ "BMI2", "1" }); + } if (ggml_cpu_has_avx512()) { features.push_back({ "AVX512", "1" }); } diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index 1832314ec1..4d4ac47c03 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -395,11 +395,11 @@ static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) { #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) -#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(RDNA2) +#if defined(CDNA) || defined(RDNA2) || defined(__gfx906__) c = __builtin_amdgcn_sdot4(a, b, c, false); #elif defined(RDNA3) c = __builtin_amdgcn_sudot4( true, a, true, b, c, false); -#elif defined(__gfx1010__) || defined(__gfx900__) +#elif defined(RDNA1) || defined(__gfx900__) int tmp1; int tmp2; asm("\n \ diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh index 46de140935..4067fd41bc 100644 --- a/ggml/src/ggml-cuda/fattn-common.cuh +++ b/ggml/src/ggml-cuda/fattn-common.cuh @@ -52,12 +52,11 @@ typedef half (*vec_dot_KQ_f16_t)( typedef float (*vec_dot_KQ_f32_t)( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8 , const void * __restrict__ Q_ds); -template +template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { const block_q4_0 * K_q4_0 = (const block_q4_0 *) K_c; - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); GGML_UNUSED(Q_v); T sum = 0.0f; @@ -93,12 +92,11 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0( return sum; } -template +template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { const block_q4_1 * K_q4_1 = (const block_q4_1 *) K_c; - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); GGML_UNUSED(Q_v); T sum = 0.0f; @@ -138,12 +136,11 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1( return sum; } -template +template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { const block_q5_0 * K_q5_0 = (const block_q5_0 *) K_c; - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); GGML_UNUSED(Q_v); T sum = 0.0f; @@ -186,12 +183,11 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0( return sum; } -template +template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { const block_q5_1 * K_q5_1 = (const block_q5_1 *) K_c; - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); GGML_UNUSED(Q_v); T sum = 0.0f; @@ -238,12 +234,11 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1( return sum; } -template +template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q8_0( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { const block_q8_0 * K_q8_0 = (const block_q8_0 *) K_c; - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); GGML_UNUSED(Q_v); T sum = 0.0f; @@ -272,12 +267,11 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q8_0( return sum; } -template +template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_f16( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8 , const void * __restrict__ Q_ds_v) { const half2 * K_h2 = (const half2 *) K_c; - constexpr int warp_size = ggml_cuda_get_physical_warp_size(); GGML_UNUSED(Q_q8); GGML_UNUSED(Q_ds_v); @@ -480,25 +474,25 @@ static __device__ __forceinline__ T dequantize_1_f16(const void * __restrict__ v return x[i]; } -template +template constexpr __device__ vec_dot_KQ_f16_t get_vec_dot_KQ_f16(ggml_type type_K) { - return type_K == GGML_TYPE_Q4_0 ? vec_dot_fattn_vec_KQ_q4_0 : - type_K == GGML_TYPE_Q4_1 ? vec_dot_fattn_vec_KQ_q4_1 : - type_K == GGML_TYPE_Q5_0 ? vec_dot_fattn_vec_KQ_q5_0 : - type_K == GGML_TYPE_Q5_1 ? vec_dot_fattn_vec_KQ_q5_1 : - type_K == GGML_TYPE_Q8_0 ? vec_dot_fattn_vec_KQ_q8_0 : - type_K == GGML_TYPE_F16 ? vec_dot_fattn_vec_KQ_f16 : + return type_K == GGML_TYPE_Q4_0 ? vec_dot_fattn_vec_KQ_q4_0 : + type_K == GGML_TYPE_Q4_1 ? vec_dot_fattn_vec_KQ_q4_1 : + type_K == GGML_TYPE_Q5_0 ? vec_dot_fattn_vec_KQ_q5_0 : + type_K == GGML_TYPE_Q5_1 ? vec_dot_fattn_vec_KQ_q5_1 : + type_K == GGML_TYPE_Q8_0 ? vec_dot_fattn_vec_KQ_q8_0 : + type_K == GGML_TYPE_F16 ? vec_dot_fattn_vec_KQ_f16 : nullptr; } -template +template constexpr __device__ vec_dot_KQ_f32_t get_vec_dot_KQ_f32(ggml_type type_K) { - return type_K == GGML_TYPE_Q4_0 ? vec_dot_fattn_vec_KQ_q4_0 : - type_K == GGML_TYPE_Q4_1 ? vec_dot_fattn_vec_KQ_q4_1 : - type_K == GGML_TYPE_Q5_0 ? vec_dot_fattn_vec_KQ_q5_0 : - type_K == GGML_TYPE_Q5_1 ? vec_dot_fattn_vec_KQ_q5_1 : - type_K == GGML_TYPE_Q8_0 ? vec_dot_fattn_vec_KQ_q8_0 : - type_K == GGML_TYPE_F16 ? vec_dot_fattn_vec_KQ_f16 : + return type_K == GGML_TYPE_Q4_0 ? vec_dot_fattn_vec_KQ_q4_0 : + type_K == GGML_TYPE_Q4_1 ? vec_dot_fattn_vec_KQ_q4_1 : + type_K == GGML_TYPE_Q5_0 ? vec_dot_fattn_vec_KQ_q5_0 : + type_K == GGML_TYPE_Q5_1 ? vec_dot_fattn_vec_KQ_q5_1 : + type_K == GGML_TYPE_Q8_0 ? vec_dot_fattn_vec_KQ_q8_0 : + type_K == GGML_TYPE_F16 ? vec_dot_fattn_vec_KQ_f16 : nullptr; } @@ -681,7 +675,8 @@ static void on_no_fattn_vec_case(const int D) { template void launch_fattn( ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, - const int nwarps, const size_t nbytes_shared, const bool need_f16_K, const bool need_f16_V + const int nwarps, const size_t nbytes_shared, const bool need_f16_K, const bool need_f16_V, + const int warp_size = WARP_SIZE ) { constexpr int ncols = ncols1 * ncols2; @@ -704,8 +699,6 @@ void launch_fattn( GGML_ASSERT(Q->ne[3] == 1); - const int warp_size = ggml_cuda_info().devices[ctx.device].warp_size; - ggml_cuda_pool & pool = ctx.pool(); cudaStream_t main_stream = ctx.stream(); const int id = ggml_cuda_get_device(); @@ -805,7 +798,6 @@ void launch_fattn( const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); GGML_ASSERT(block_dim.x % warp_size == 0); - GGML_ASSERT(!GGML_CUDA_CC_IS_AMD(cc) || block_dim.x * block_dim.y <= 4 * (unsigned int)warp_size); fattn_kernel<<>>( (const char *) Q->data, K_data, diff --git a/ggml/src/ggml-cuda/fattn-wmma-f16.cu b/ggml/src/ggml-cuda/fattn-wmma-f16.cu index 622cf28576..dab1d5cbca 100644 --- a/ggml/src/ggml-cuda/fattn-wmma-f16.cu +++ b/ggml/src/ggml-cuda/fattn-wmma-f16.cu @@ -469,6 +469,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggm constexpr int frag_m = cols_per_block == 8 && D % 32 == 0 ? 32 : 16; const int blocks_num_pb1 = ((Q->ne[1] + cols_per_block - 1) / cols_per_block)*Q->ne[2]*Q->ne[3]; const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm; + const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size; float logit_softcap; memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float)); @@ -485,7 +486,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggm fattn_kernel = flash_attn_ext_f16< D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>; } - launch_fattn(ctx, dst, fattn_kernel, nwarps, 0, true, true); + launch_fattn(ctx, dst, fattn_kernel, nwarps, 0, true, true, warp_size); return; } if (2*blocks_num_pb1 < 2*nsm) { @@ -500,7 +501,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggm fattn_kernel = flash_attn_ext_f16< D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>; } - launch_fattn(ctx, dst, fattn_kernel, nwarps, 0, true, true); + launch_fattn(ctx, dst, fattn_kernel, nwarps, 0, true, true, warp_size); return; } constexpr int parallel_blocks = 1; @@ -514,7 +515,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggm fattn_kernel = flash_attn_ext_f16< D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>; } - launch_fattn(ctx, dst, fattn_kernel, nwarps, 0, true, true); + launch_fattn(ctx, dst, fattn_kernel, nwarps, 0, true, true, warp_size); } void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu index 24f973056a..2e72fc8fd3 100644 --- a/ggml/src/ggml-cuda/fattn.cu +++ b/ggml/src/ggml-cuda/fattn.cu @@ -310,7 +310,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst } // The MMA implementation needs Turing or newer, use the old WMMA code for Volta: - if (cc == GGML_CUDA_CC_VOLTA) { + if (fp16_mma_available(cc) && !new_mma_available(cc)) { ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst); return; } diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index b5d2c84111..497de37be8 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -2571,7 +2571,7 @@ static void maintain_cuda_graph(ggml_backend_cuda_context * cuda_ctx, std::vecto for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) { if(count(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), cuda_ctx->cuda_graph->params[i].func) > 0) { char ** updated_kernel_arg_ptr = cuda_ctx->cuda_graph->updated_kernel_arg.at(k++); - cuda_ctx->cuda_graph->params[i].kernelParams[1] = updated_kernel_arg_ptr; + *(void**)cuda_ctx->cuda_graph->params[i].kernelParams[1] = *(void**)updated_kernel_arg_ptr; CUDA_CHECK(cudaGraphKernelNodeSetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i])); } } diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu index 4fb466ca00..a7d518a574 100644 --- a/ggml/src/ggml-cuda/mmvq.cu +++ b/ggml/src/ggml-cuda/mmvq.cu @@ -47,11 +47,89 @@ static constexpr __device__ int get_vdr_mmvq(ggml_type type) { 1; } +enum mmvq_parameter_table_id { + MMVQ_PARAMETERS_GENERIC = 0, + MMVQ_PARAMETERS_GCN, + MMVQ_PARAMETERS_RDNA2 +}; + +static constexpr __device__ mmvq_parameter_table_id get_device_table_id() { +#if defined(RDNA2) || defined(RDNA3) + return MMVQ_PARAMETERS_RDNA2; +#elif defined(GCN) || defined(CDNA) + return MMVQ_PARAMETERS_GCN; +#else + return MMVQ_PARAMETERS_GENERIC; +#endif +} + +static __host__ mmvq_parameter_table_id get_device_table_id(int cc) { + if (GGML_CUDA_CC_IS_RDNA2(cc) || GGML_CUDA_CC_IS_RDNA3(cc)) { + return MMVQ_PARAMETERS_RDNA2; + } + if (GGML_CUDA_CC_IS_GCN(cc) || GGML_CUDA_CC_IS_CDNA(cc)) { + return MMVQ_PARAMETERS_GCN; + } + return MMVQ_PARAMETERS_GENERIC; +} + +static constexpr __host__ __device__ int calc_nwarps(int ncols_y, mmvq_parameter_table_id table_id) { + if (table_id == MMVQ_PARAMETERS_GENERIC) { + switch (ncols_y) { + case 1: + case 2: + case 3: + case 4: + return 4; + case 5: + case 6: + case 7: + case 8: + return 2; + default: + return 1; + } + } else if (table_id == MMVQ_PARAMETERS_GCN) { + switch (ncols_y) { + case 1: + case 2: + case 3: + case 4: + return 2; + case 5: + case 6: + case 7: + case 8: + default: + return 1; + } + } + return 1; +} + +static constexpr __host__ __device__ int calc_rows_per_block(int ncols_y, int table_id) { + if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN) { + switch (ncols_y) { + case 1: + return 1; + case 2: + case 3: + case 4: + case 5: + case 6: + case 7: + case 8: + return 2; + default: + return 1; + } + } + return 1; +} + template -#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) // tell the compiler to use as many registers as it wants, see nwarps definition below -__launch_bounds__((ncols_y <= 4 ? 4 : 2)*WARP_SIZE, 1) -#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) +__launch_bounds__(calc_nwarps(ncols_y, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1) static __global__ void mul_mat_vec_q( const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) { @@ -59,24 +137,20 @@ static __global__ void mul_mat_vec_q( constexpr int qk = ggml_cuda_type_traits::qk; constexpr int qi = ggml_cuda_type_traits::qi; constexpr int vdr = get_vdr_mmvq(type); + constexpr mmvq_parameter_table_id table_id = get_device_table_id(); + constexpr int nwarps = calc_nwarps(ncols_y, table_id); + constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_y, table_id); + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type); -#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3)) - constexpr int nwarps = 1; - constexpr int rows_per_cuda_block = 1; -#else - constexpr int nwarps = ncols_y <= 4 ? 4 : 2; - constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2; -#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3) - - const int tid = WARP_SIZE*threadIdx.y + threadIdx.x; + const int tid = warp_size*threadIdx.y + threadIdx.x; const int row0 = rows_per_cuda_block*blockIdx.x; const int blocks_per_row_x = ncols_x / qk; const int blocks_per_col_y = nrows_y / QK8_1; - constexpr int blocks_per_iter = vdr * nwarps*WARP_SIZE / qi; + constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi; -// partial sum for each thread + // partial sum for each thread float tmp[ncols_y][rows_per_cuda_block] = {0.0f}; const block_q8_1 * y = (const block_q8_1 *) vy; @@ -96,7 +170,7 @@ static __global__ void mul_mat_vec_q( } } - __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][WARP_SIZE]; + __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][warp_size]; if (threadIdx.y > 0) { #pragma unroll for (int j = 0; j < ncols_y; ++j) { @@ -120,7 +194,7 @@ static __global__ void mul_mat_vec_q( for (int l = 0; l < nwarps-1; ++l) { tmp[j][i] += tmp_shared[l][j][i][threadIdx.x]; } - tmp[j][i] = warp_reduce_sum(tmp[j][i]); + tmp[j][i] = warp_reduce_sum(tmp[j][i]); } if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + threadIdx.x < nrows_dst)) { @@ -129,6 +203,13 @@ static __global__ void mul_mat_vec_q( } } +static std::pair calc_launch_params(const int ncols_y, const int nrows_x, const int warp_size, const mmvq_parameter_table_id table_id) { + const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_y, table_id) - 1) / calc_rows_per_block(ncols_y, table_id); + const dim3 block_nums(nblocks, 1, 1); + const dim3 block_dims(warp_size, calc_nwarps(ncols_y, table_id), 1); + return {block_nums, block_dims}; +} + template static void mul_mat_vec_q_cuda( const void * vx, const void * vy, float * dst, @@ -137,65 +218,67 @@ static void mul_mat_vec_q_cuda( GGML_ASSERT(ncols_x % ggml_blck_size(type) == 0); GGML_ASSERT(ncols_y <= MMVQ_MAX_BATCH_SIZE); - int id = ggml_cuda_get_device(); - - int64_t nwarps = 1; - int64_t rows_per_cuda_block = 1; - - if (ggml_cuda_info().devices[id].cc < GGML_CUDA_CC_RDNA2) { // NVIDIA and AMD older than RDNA2 - switch(ncols_y) { - case 1: - nwarps = 4; - rows_per_cuda_block = 1; - break; - case 2: - case 3: - case 4: - nwarps = 4; - rows_per_cuda_block = 2; - break; - case 5: - case 6: - case 7: - case 8: - nwarps = 2; - rows_per_cuda_block = 2; - break; - default: - GGML_ABORT("fatal error"); - break; - } - } - - const int64_t nblocks = (nrows_x + rows_per_cuda_block - 1) / rows_per_cuda_block; - const dim3 block_nums(nblocks, 1, 1); - const dim3 block_dims(WARP_SIZE, nwarps, 1); + const int device = ggml_cuda_get_device(); + const int warp_size = ggml_cuda_info().devices[device].warp_size; + const mmvq_parameter_table_id table_id = get_device_table_id(ggml_cuda_info().devices[device].cc); switch (ncols_y) { case 1: - mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + { + constexpr int c_ncols_y = 1; + std::pair dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id); + mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); break; + } case 2: - mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + { + constexpr int c_ncols_y = 2; + std::pair dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id); + mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); break; + } case 3: - mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + { + constexpr int c_ncols_y = 3; + std::pair dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id); + mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); break; + } case 4: - mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + { + constexpr int c_ncols_y = 4; + std::pair dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id); + mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); break; + } case 5: - mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + { + constexpr int c_ncols_y = 5; + std::pair dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id); + mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); break; + } case 6: - mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + { + constexpr int c_ncols_y = 6; + std::pair dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id); + mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); break; + } case 7: - mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + { + constexpr int c_ncols_y = 7; + std::pair dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id); + mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); break; + } case 8: - mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + { + constexpr int c_ncols_y = 8; + std::pair dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id); + mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); break; + } default: GGML_ABORT("fatal error"); break; diff --git a/ggml/src/ggml-metal/CMakeLists.txt b/ggml/src/ggml-metal/CMakeLists.txt index 89fcde2faa..e222327809 100644 --- a/ggml/src/ggml-metal/CMakeLists.txt +++ b/ggml/src/ggml-metal/CMakeLists.txt @@ -27,12 +27,12 @@ configure_file(../ggml-common.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY) configure_file(ggml-metal-impl.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h COPYONLY) +set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/../ggml-common.h") if (GGML_METAL_EMBED_LIBRARY) enable_language(ASM) add_compile_definitions(GGML_METAL_EMBED_LIBRARY) - set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/../ggml-common.h") set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal") set(METALLIB_IMPL "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h") @@ -88,12 +88,11 @@ else() add_custom_command( OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib - COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air - COMMAND xcrun -sdk macosx metallib ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib - COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air + COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o - | + xcrun -sdk macosx metallib - -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal - DEPENDS ggml-metal.metal ggml-common.h + DEPENDS ggml-metal.metal ${METALLIB_COMMON} COMMENT "Compiling Metal kernels" ) diff --git a/ggml/src/ggml-metal/ggml-metal-impl.h b/ggml/src/ggml-metal/ggml-metal-impl.h index e3dc25f168..a58c474eb0 100644 --- a/ggml/src/ggml-metal/ggml-metal-impl.h +++ b/ggml/src/ggml-metal/ggml-metal-impl.h @@ -285,4 +285,239 @@ typedef struct { float eps; } ggml_metal_kargs_rms_norm; +typedef struct { + int64_t ne00; + int64_t ne01; + int64_t ne02; + uint64_t nb00; + uint64_t nb01; + uint64_t nb02; + int32_t n_groups; + float eps; +} ggml_metal_kargs_group_norm; + +typedef struct { + int32_t IC; + int32_t IL; + int32_t K; + int32_t s0; + uint64_t nb0; + uint64_t nb1; +} ggml_metal_kargs_conv_transpose_1d; + +typedef struct { + uint64_t ofs0; + uint64_t ofs1; + int32_t IW; + int32_t IH; + int32_t CHW; + int32_t s0; + int32_t s1; + int32_t p0; + int32_t p1; + int32_t d0; + int32_t d1; + int32_t N; + int32_t KH; + int32_t KW; + int32_t KHW; // KH * KW, pre-computed on CPU to save GPU resources +} ggml_metal_kargs_im2col; + +typedef struct { + int64_t ne00; + int64_t ne01; + int64_t ne02; + int64_t ne03; + uint64_t nb00; + uint64_t nb01; + uint64_t nb02; + uint64_t nb03; + int64_t ne10; + int64_t ne11; + int64_t ne12; + int64_t ne13; + uint64_t nb10; + uint64_t nb11; + uint64_t nb12; + uint64_t nb13; + int64_t ne0; + int64_t ne1; + int64_t ne2; + int64_t ne3; + uint64_t nb0; + uint64_t nb1; + uint64_t nb2; + uint64_t nb3; +} ggml_metal_kargs_sum_rows; + +typedef struct { + int64_t ne00; + int64_t ne01; + int64_t ne02; + float scale; + float max_bias; + float m0; + float m1; + uint32_t n_head_log2; +} ggml_metal_kargs_soft_max; + +typedef struct { + int64_t ne00; + int64_t ne01; + int n_past; +} ggml_metal_kargs_diag_mask_inf; + +typedef struct { + int64_t ne00; + int64_t ne01; + int64_t ne02; + uint64_t nb00; + uint64_t nb01; + uint64_t nb02; + int64_t ne10; + int64_t ne11; + uint64_t nb10; + uint64_t nb11; + int64_t ne0; + int64_t ne1; + int64_t ne2; + uint64_t nb0; + uint64_t nb1; + uint64_t nb2; +} ggml_metal_kargs_ssm_conv; + +typedef struct { + int64_t d_state; + int64_t d_inner; + int64_t n_seq_tokens; + int64_t n_seqs; + uint64_t nb00; + uint64_t nb01; + uint64_t nb02; + uint64_t nb10; + uint64_t nb11; + uint64_t nb12; + uint64_t nb13; + uint64_t nb20; + uint64_t nb21; + uint64_t nb22; + uint64_t nb30; + uint64_t nb31; + uint64_t nb40; + uint64_t nb41; + uint64_t nb42; + uint64_t nb50; + uint64_t nb51; + uint64_t nb52; +} ggml_metal_kargs_ssm_scan; + +typedef struct { + int64_t ne00; + uint64_t nb01; + uint64_t nb02; + int64_t ne10; + uint64_t nb10; + uint64_t nb11; + uint64_t nb1; + uint64_t nb2; +} ggml_metal_kargs_get_rows; + +typedef struct { + int64_t ne00; + int64_t ne01; + int64_t ne02; + int64_t ne03; + uint64_t nb00; + uint64_t nb01; + uint64_t nb02; + uint64_t nb03; + int64_t ne0; + int64_t ne1; + int64_t ne2; + int64_t ne3; + uint64_t nb0; + uint64_t nb1; + uint64_t nb2; + uint64_t nb3; + float sf0; + float sf1; + float sf2; + float sf3; +} ggml_metal_kargs_upscale; + +typedef struct { + int64_t ne00; + int64_t ne01; + int64_t ne02; + int64_t ne03; + uint64_t nb00; + uint64_t nb01; + uint64_t nb02; + uint64_t nb03; + int64_t ne0; + int64_t ne1; + int64_t ne2; + int64_t ne3; + uint64_t nb0; + uint64_t nb1; + uint64_t nb2; + uint64_t nb3; +} ggml_metal_kargs_pad; + +typedef struct { + int64_t ne00; + int64_t ne01; + int64_t ne02; + int64_t ne03; + uint64_t nb00; + uint64_t nb01; + uint64_t nb02; + uint64_t nb03; + int64_t ne0; + int64_t ne1; + int64_t ne2; + int64_t ne3; + uint64_t nb0; + uint64_t nb1; + uint64_t nb2; + uint64_t nb3; + int32_t p0; + int32_t p1; +} ggml_metal_kargs_pad_reflect_1d; + +typedef struct { + uint64_t nb1; + int dim; + int max_period; +} ggml_metal_kargs_timestep_embedding; + +typedef struct { + float slope; +} ggml_metal_kargs_leaky_relu; + +typedef struct { + int64_t ncols; + int64_t ncols_pad; +} ggml_metal_kargs_argsort; + +typedef struct { + int64_t ne0; + float start; + float step; +} ggml_metal_kargs_arange; + +typedef struct { + int32_t k0; + int32_t k1; + int32_t s0; + int32_t s1; + int32_t p0; + int32_t p1; + int64_t IH; + int64_t IW; + int64_t OH; + int64_t OW; + int64_t parallel_elements; +} ggml_metal_kargs_pool_2d; + #endif // GGML_METAL_IMPL diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m index 9eb9a0db68..e51a4169a2 100644 --- a/ggml/src/ggml-metal/ggml-metal.m +++ b/ggml/src/ggml-metal/ggml-metal.m @@ -46,6 +46,7 @@ static struct ggml_backend_device g_ggml_backend_metal_device; static struct ggml_backend_metal_device_context { id mtl_device; int mtl_device_ref_count; + id mtl_library; bool has_simdgroup_reduction; bool has_simdgroup_mm; @@ -57,6 +58,7 @@ static struct ggml_backend_metal_device_context { } g_ggml_ctx_dev_main = { /*.mtl_device =*/ nil, /*.mtl_device_ref_count =*/ 0, + /*.mtl_library =*/ nil, /*.has_simdgroup_reduction =*/ false, /*.has_simdgroup_mm =*/ false, /*.has_residency_sets =*/ false, @@ -108,6 +110,11 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte ctx->mtl_device_ref_count--; if (ctx->mtl_device_ref_count == 0) { + if (ctx->mtl_library) { + [ctx->mtl_library release]; + ctx->mtl_library = nil; + } + if (ctx->mtl_device) { [ctx->mtl_device release]; ctx->mtl_device = nil; @@ -467,11 +474,13 @@ struct ggml_backend_metal_context { // for now it is easier to work in a separate file // static NSString * const msl_library_source = @"see metal.metal"; +#if !GGML_METAL_EMBED_LIBRARY // Here to assist with NSBundle Path Hack @interface GGMLMetalClass : NSObject @end @implementation GGMLMetalClass @end +#endif static void * ggml_metal_host_malloc(size_t n) { void * data = NULL; @@ -493,6 +502,139 @@ static void * ggml_metal_host_malloc(size_t n) { return data; } +// load library +// +// - first check if the library is embedded +// - then check if the library is in the bundle +// - if not found, load the source and compile it +// - if that fails, return NULL +static id ggml_metal_load_library(id device, bool use_bfloat) { + id metal_library = nil; + NSError * error = nil; + NSString * src = nil; + +#if GGML_METAL_EMBED_LIBRARY + GGML_LOG_INFO("%s: using embedded metal library\n", __func__); + + extern const char ggml_metallib_start[]; + extern const char ggml_metallib_end[]; + + src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding]; + +#else + +#ifdef SWIFT_PACKAGE + NSBundle * bundle = SWIFTPM_MODULE_BUNDLE; +#else + NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]]; +#endif + + NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"]; + if (path_lib == nil) { + // Try to find the resource in the directory where the current binary located. + NSString * current_binary = [[NSProcessInfo processInfo] arguments][0]; + NSString * bin_dir = [current_binary stringByDeletingLastPathComponent]; + NSString * default_metallib_path = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]]; + if ([[NSFileManager defaultManager] isReadableFileAtPath:default_metallib_path]) { + GGML_LOG_INFO("%s: found '%s'\n", __func__, [default_metallib_path UTF8String]); + NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:default_metallib_path error:&error]; + if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) { + // Optionally, if this is a symlink, try to resolve it. + default_metallib_path = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:default_metallib_path error:&error]; + if (default_metallib_path && [default_metallib_path length] > 0 && ![[default_metallib_path substringToIndex:1] isEqualToString:@"/"]) { + // It is a relative path, adding the binary directory as directory prefix. + default_metallib_path = [NSString pathWithComponents:@[bin_dir, default_metallib_path]]; + } + if (!default_metallib_path || ![[NSFileManager defaultManager] isReadableFileAtPath:default_metallib_path]) { + // Link to the resource could not be resolved. + default_metallib_path = nil; + } else { + GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [default_metallib_path UTF8String]); + } + } + } else { + // The resource couldn't be found in the binary's directory. + default_metallib_path = nil; + } + path_lib = default_metallib_path; + } + + if (path_lib != nil) { + // pre-compiled library found + NSURL * libURL = [NSURL fileURLWithPath:path_lib]; + GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]); + + metal_library = [device newLibraryWithURL:libURL error:&error]; + if (error) { + GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); + return NULL; + } + } else { + GGML_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__); + + NSString * path_source; + NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"]; + + GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil"); + + if (path_resource) { + path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"]; + } else { + path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"]; + } + + if (path_source == nil) { + GGML_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__); + path_source = @"ggml-metal.metal"; + } + + GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]); + + src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error]; + if (error) { + GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); + return NULL; + } + } +#endif + + if (!metal_library) { + @autoreleasepool { + // dictionary of preprocessor macros + NSMutableDictionary * prep = [NSMutableDictionary dictionary]; + + if (use_bfloat) { + [prep setObject:@"1" forKey:@"GGML_METAL_USE_BF16"]; + } + +#if GGML_METAL_EMBED_LIBRARY + [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"]; +#endif + + MTLCompileOptions * options = [MTLCompileOptions new]; + options.preprocessorMacros = prep; + + //[options setFastMathEnabled:false]; + + metal_library = [device newLibraryWithSource:src options:options error:&error]; + if (error) { + GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); + return NULL; + } + +#if !__has_feature(objc_arc) + [options release]; +#endif + } + } + +#if GGML_METAL_EMBED_LIBRARY + [src release]; +#endif // GGML_METAL_EMBED_LIBRARY + + return metal_library; +} + static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t dev) { GGML_LOG_INFO("%s: allocating\n", __func__); @@ -520,137 +662,14 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT); - id metal_library; - // load library - // - // - first check if the library is embedded - // - then check if the library is in the bundle - // - if not found, load the source and compile it - // - if that fails, return NULL - { - NSBundle * bundle = nil; -#ifdef SWIFT_PACKAGE - bundle = SWIFTPM_MODULE_BUNDLE; -#else - bundle = [NSBundle bundleForClass:[GGMLMetalClass class]]; -#endif - - NSError * error = nil; - -#if GGML_METAL_EMBED_LIBRARY - const bool try_metallib = false; -#else - const bool try_metallib = true; -#endif - - NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"]; - if (path_lib == nil) { - // Try to find the resource in the directory where the current binary located. - NSString * current_binary = [[NSProcessInfo processInfo] arguments][0]; - NSString * bin_dir = [current_binary stringByDeletingLastPathComponent]; - NSString * default_metallib_path = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]]; - if ([[NSFileManager defaultManager] isReadableFileAtPath:default_metallib_path]) { - GGML_LOG_INFO("%s: found '%s'\n", __func__, [default_metallib_path UTF8String]); - NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:default_metallib_path error:&error]; - if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) { - // Optionally, if this is a symlink, try to resolve it. - default_metallib_path = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:default_metallib_path error:&error]; - if (default_metallib_path && [default_metallib_path length] > 0 && ![[default_metallib_path substringToIndex:1] isEqualToString:@"/"]) { - // It is a relative path, adding the binary directory as directory prefix. - default_metallib_path = [NSString pathWithComponents:@[bin_dir, default_metallib_path]]; - } - if (!default_metallib_path || ![[NSFileManager defaultManager] isReadableFileAtPath:default_metallib_path]) { - // Link to the resource could not be resolved. - default_metallib_path = nil; - } else { - GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [default_metallib_path UTF8String]); - } - } - } else { - // The resource couldn't be found in the binary's directory. - default_metallib_path = nil; - } - path_lib = default_metallib_path; - } - - if (try_metallib && path_lib != nil) { - // pre-compiled library found - NSURL * libURL = [NSURL fileURLWithPath:path_lib]; - GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]); - - metal_library = [device newLibraryWithURL:libURL error:&error]; - if (error) { - GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); - return NULL; - } - } else { -#if GGML_METAL_EMBED_LIBRARY - GGML_LOG_INFO("%s: using embedded metal library\n", __func__); - - extern const char ggml_metallib_start[]; - extern const char ggml_metallib_end[]; - - NSString * src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding]; -#else - GGML_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__); - - NSString * path_source; - NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"]; - - GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil"); - - if (path_resource) { - path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"]; - } else { - path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"]; - } - - if (path_source == nil) { - GGML_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__); - path_source = @"ggml-metal.metal"; - } - - GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]); - - NSString * src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error]; - if (error) { - GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); - return NULL; - } -#endif // GGML_METAL_EMBED_LIBRARY - - @autoreleasepool { - // dictionary of preprocessor macros - NSMutableDictionary * prep = [NSMutableDictionary dictionary]; - - if (ctx_dev->use_bfloat) { - [prep setObject:@"1" forKey:@"GGML_METAL_USE_BF16"]; - } - -#if GGML_METAL_EMBED_LIBRARY - [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"]; -#endif - - MTLCompileOptions * options = [MTLCompileOptions new]; - options.preprocessorMacros = prep; - - //[options setFastMathEnabled:false]; - - metal_library = [device newLibraryWithSource:src options:options error:&error]; - if (error) { - GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); - return NULL; - } - -#if !__has_feature(objc_arc) - [options release]; -#endif - } -#if GGML_METAL_EMBED_LIBRARY - [src release]; -#endif // GGML_METAL_EMBED_LIBRARY - } + if (ctx_dev->mtl_library == nil) { + ctx_dev->mtl_library = ggml_metal_load_library(device, ctx_dev->use_bfloat); + } + id metal_library = ctx_dev->mtl_library; + if (metal_library == nil) { + GGML_LOG_ERROR("%s: error: metal library is nil\n", __func__); + return NULL; } // print MTL GPU family: @@ -724,7 +743,6 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de [metal_function release]; \ if (error) { \ GGML_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \ - [metal_library release]; \ return NULL; \ } \ } else { \ @@ -1043,8 +1061,6 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32, pool_2d_max_f32, true); } - [metal_library release]; - return ctx; } @@ -1944,34 +1960,38 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline; - // TODO: add ggml_metal_kargs struct + + ggml_metal_kargs_sum_rows args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.ne02 =*/ ne02, + /*.ne03 =*/ ne03, + /*.nb00 =*/ nb00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.nb03 =*/ nb03, + /*.ne10 =*/ ne10, + /*.ne11 =*/ ne11, + /*.ne12 =*/ ne12, + /*.ne13 =*/ ne13, + /*.nb10 =*/ nb10, + /*.nb11 =*/ nb11, + /*.nb12 =*/ nb12, + /*.nb13 =*/ nb13, + /*.ne0 =*/ ne0, + /*.ne1 =*/ ne1, + /*.ne2 =*/ ne2, + /*.ne3 =*/ ne3, + /*.nb0 =*/ nb0, + /*.nb1 =*/ nb1, + /*.nb2 =*/ nb2, + /*.nb3 =*/ nb3, + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; - [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; - [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9]; - [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:10]; - [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11]; - [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12]; - [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:13]; - [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14]; - [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15]; - [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16]; - [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:18]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:19]; - [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:20]; - [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:21]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:22]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:23]; - [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:24]; - [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:25]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; } break; @@ -2020,8 +2040,17 @@ static void ggml_metal_encode_node( const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); - // TODO: add ggml_metal_kargs struct - // TODO: optimize (see https://github.com/ggml-org/llama.cpp/pull/10238/commits/7941b6b9ec29a2866fec6fa6c51612515ca509f6) + ggml_metal_kargs_soft_max args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.ne02 =*/ ne02, + /*.scale =*/ scale, + /*.max_bias =*/ max_bias, + /*.m0 =*/ m0, + /*.m1 =*/ m1, + /*.n_head_log2 =*/ n_head_log2, + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; if (id_src1) { @@ -2030,14 +2059,7 @@ static void ggml_metal_encode_node( [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1]; } [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5]; - [encoder setBytes:&scale length:sizeof(scale) atIndex:6]; - [encoder setBytes:&max_bias length:sizeof(max_bias) atIndex:7]; - [encoder setBytes:&m0 length:sizeof(m0) atIndex:8]; - [encoder setBytes:&m1 length:sizeof(m1) atIndex:9]; - [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:10]; + [encoder setBytes:&args length:sizeof(args) atIndex:3]; [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0]; @@ -2055,13 +2077,16 @@ static void ggml_metal_encode_node( pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF].pipeline; } - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_diag_mask_inf args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.n_past =*/ n_past, + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; - [encoder setBytes:&n_past length:sizeof(int) atIndex:4]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; if (ne00%8 == 0) { [encoder dispatchThreadgroups:MTLSizeMake(ne00*ne01*ne02/8, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; @@ -2080,27 +2105,30 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_CONV_F32].pipeline; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_ssm_conv args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.ne02 =*/ ne02, + /*.nb00 =*/ nb00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.ne10 =*/ ne10, + /*.ne11 =*/ ne11, + /*.nb10 =*/ nb10, + /*.nb11 =*/ nb11, + /*.ne0 =*/ ne0, + /*.ne1 =*/ ne1, + /*.ne2 =*/ ne2, + /*.nb0 =*/ nb0, + /*.nb1 =*/ nb1, + /*.nb2 =*/ nb2, + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; - [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9]; - [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10]; - [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:11]; - [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:12]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:13]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:14]; - [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:15]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:16]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:17]; - [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:18]; + [encoder setBytes:&args length:sizeof(args) atIndex:3]; [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne1, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; } break; @@ -2151,7 +2179,31 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_ssm_scan args = { + /*.d_state =*/ d_state, + /*.d_inner =*/ d_inner, + /*.n_seq_tokens =*/ n_seq_tokens, + /*.n_seqs =*/ n_seqs, + /*.nb00 =*/ nb00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.nb10 =*/ nb10, + /*.nb11 =*/ nb11, + /*.nb12 =*/ nb12, + /*.nb13 =*/ nb13, + /*.nb20 =*/ nb20, + /*.nb21 =*/ nb21, + /*.nb22 =*/ nb22, + /*.nb30 =*/ nb30, + /*.nb31 =*/ nb31, + /*.nb40 =*/ nb40, + /*.nb41 =*/ nb41, + /*.nb42 =*/ nb42, + /*.nb50 =*/ nb50, + /*.nb51 =*/ nb51, + /*.nb52 =*/ nb52, + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; @@ -2160,30 +2212,7 @@ static void ggml_metal_encode_node( [encoder setBuffer:id_src4 offset:offs_src4 atIndex:4]; [encoder setBuffer:id_src5 offset:offs_src5 atIndex:5]; [encoder setBuffer:id_dst offset:offs_dst atIndex:6]; - - [encoder setBytes:&d_state length:sizeof(d_state) atIndex:7]; - [encoder setBytes:&d_inner length:sizeof(d_inner) atIndex:8]; - [encoder setBytes:&n_seq_tokens length:sizeof(n_seq_tokens) atIndex:9]; - [encoder setBytes:&n_seqs length:sizeof(n_seqs) atIndex:10]; - - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:11]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:12]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:13]; - [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14]; - [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15]; - [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16]; - [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17]; - [encoder setBytes:&nb20 length:sizeof(nb20) atIndex:18]; - [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:19]; - [encoder setBytes:&nb22 length:sizeof(nb22) atIndex:20]; - [encoder setBytes:&nb30 length:sizeof(nb30) atIndex:21]; - [encoder setBytes:&nb31 length:sizeof(nb31) atIndex:22]; - [encoder setBytes:&nb40 length:sizeof(nb40) atIndex:23]; - [encoder setBytes:&nb41 length:sizeof(nb41) atIndex:24]; - [encoder setBytes:&nb42 length:sizeof(nb42) atIndex:25]; - [encoder setBytes:&nb50 length:sizeof(nb50) atIndex:26]; - [encoder setBytes:&nb51 length:sizeof(nb51) atIndex:27]; - [encoder setBytes:&nb52 length:sizeof(nb52) atIndex:28]; + [encoder setBytes:&args length:sizeof(args) atIndex:7]; [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; } break; @@ -3040,19 +3069,22 @@ static void ggml_metal_encode_node( default: GGML_ABORT("not implemented"); } - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_get_rows args = { + /*.ne00 =*/ ne00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.ne10 =*/ ne10, + /*.nb10 =*/ nb10, + /*.nb11 =*/ nb11, + /*.nb1 =*/ nb1, + /*.nb2 =*/ nb2, + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:3]; - [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:4]; - [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:5]; - [encoder setBytes:&ne10 length:sizeof( int64_t) atIndex:6]; - [encoder setBytes:&nb10 length:sizeof( int64_t) atIndex:7]; - [encoder setBytes:&nb11 length:sizeof( int64_t) atIndex:8]; - [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:9]; - [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:10]; + [encoder setBytes:&args length:sizeof(args) atIndex:3]; [encoder dispatchThreadgroups:MTLSizeMake(ne10, ne11, 1) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)]; } break; @@ -3109,18 +3141,21 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GROUP_NORM].pipeline; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_group_norm args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.ne02 =*/ ne02, + /*.nb00 =*/ nb00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.n_groups =*/ n_groups, + /*.eps =*/ eps, + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4]; - [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:5]; - [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:6]; - [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:7]; - [encoder setBytes:&n_groups length:sizeof( int32_t) atIndex:8]; - [encoder setBytes:&eps length:sizeof( float) atIndex:9]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0]; [encoder dispatchThreadgroups:MTLSizeMake(n_groups, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; @@ -3278,8 +3313,8 @@ static void ggml_metal_encode_node( const int32_t CHW = IC * KH * KW; - const int32_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4; - const int32_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4; + const uint64_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4; + const uint64_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4; id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F32].pipeline; @@ -3301,27 +3336,30 @@ static void ggml_metal_encode_node( default: GGML_ABORT("fatal error"); }; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_im2col args = { + /*.ofs0 =*/ ofs0, + /*.ofs1 =*/ ofs1, + /*.IW =*/ IW, + /*.IH =*/ IH, + /*.CHW =*/ CHW, + /*.s0 =*/ s0, + /*.s1 =*/ s1, + /*.p0 =*/ p0, + /*.p1 =*/ p1, + /*.d0 =*/ d0, + /*.d1 =*/ d1, + /*.N =*/ N, + /*.KH =*/ KH, + /*.KW =*/ KW, + /*.KHW =*/ KH * KW, + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src1 offset:offs_src1 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ofs0 length:sizeof(int32_t) atIndex:2]; - [encoder setBytes:&ofs1 length:sizeof(int32_t) atIndex:3]; - [encoder setBytes:&IW length:sizeof(int32_t) atIndex:4]; - [encoder setBytes:&IH length:sizeof(int32_t) atIndex:5]; - [encoder setBytes:&CHW length:sizeof(int32_t) atIndex:6]; - [encoder setBytes:&s0 length:sizeof(int32_t) atIndex:7]; - [encoder setBytes:&s1 length:sizeof(int32_t) atIndex:8]; - [encoder setBytes:&p0 length:sizeof(int32_t) atIndex:9]; - [encoder setBytes:&p1 length:sizeof(int32_t) atIndex:10]; - [encoder setBytes:&d0 length:sizeof(int32_t) atIndex:11]; - [encoder setBytes:&d1 length:sizeof(int32_t) atIndex:12]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; if (is_gt_mttpt) { - [encoder setBytes:&N length:sizeof(int32_t) atIndex:13]; - [encoder setBytes:&KH length:sizeof(int32_t) atIndex:14]; - [encoder setBytes:&KW length:sizeof(int32_t) atIndex:15]; - const uint64_t n_threads = MIN(pipeline.maxTotalThreadsPerThreadgroup, (uint64_t)N); const int64_t quotient = N / n_threads + (N % n_threads > 0 ? 1 : 0); @@ -3361,16 +3399,20 @@ static void ggml_metal_encode_node( default: GGML_ABORT("fatal error"); }; + ggml_metal_kargs_conv_transpose_1d args = { + /*.IC =*/ IC, + /*.IL =*/ IL, + /*.K =*/ K, + /*.s0 =*/ s0, + /*.nb0 =*/ nb0, + /*.nb1 =*/ nb1, + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&IC length:sizeof( int32_t) atIndex:3]; - [encoder setBytes:&IL length:sizeof( int32_t) atIndex:4]; - [encoder setBytes:&K length:sizeof( int32_t) atIndex:5]; - [encoder setBytes:&s0 length:sizeof( int32_t) atIndex:6]; - [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:7]; - [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:8]; + [encoder setBytes:&args length:sizeof(args) atIndex:3]; [encoder dispatchThreadgroups:MTLSizeMake(OL, OC, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; } break; @@ -3385,30 +3427,33 @@ static void ggml_metal_encode_node( const id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_UPSCALE_F32].pipeline; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_upscale args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.ne02 =*/ ne02, + /*.ne03 =*/ ne03, + /*.nb00 =*/ nb00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.nb03 =*/ nb03, + /*.ne0 =*/ ne0, + /*.ne1 =*/ ne1, + /*.ne2 =*/ ne2, + /*.ne3 =*/ ne3, + /*.nb0 =*/ nb0, + /*.nb1 =*/ nb1, + /*.nb2 =*/ nb2, + /*.nb3 =*/ nb3, + /*.sf0 =*/ sf0, + /*.sf1 =*/ sf1, + /*.sf2 =*/ sf2, + /*.sf3 =*/ sf3 + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; - [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; - [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11]; - [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12]; - [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15]; - [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16]; - [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17]; - [encoder setBytes:&sf0 length:sizeof(sf0) atIndex:18]; - [encoder setBytes:&sf1 length:sizeof(sf1) atIndex:19]; - [encoder setBytes:&sf2 length:sizeof(sf2) atIndex:20]; - [encoder setBytes:&sf3 length:sizeof(sf3) atIndex:21]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0); @@ -3420,26 +3465,29 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_PAD_F32].pipeline; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_pad args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.ne02 =*/ ne02, + /*.ne03 =*/ ne03, + /*.nb00 =*/ nb00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.nb03 =*/ nb03, + /*.ne0 =*/ ne0, + /*.ne1 =*/ ne1, + /*.ne2 =*/ ne2, + /*.ne3 =*/ ne3, + /*.nb0 =*/ nb0, + /*.nb1 =*/ nb1, + /*.nb2 =*/ nb2, + /*.nb3 =*/ nb3 + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; - [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; - [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11]; - [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12]; - [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15]; - [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16]; - [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; const int nth = MIN(1024, ne0); @@ -3454,24 +3502,31 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_PAD_REFLECT_1D_F32].pipeline; + ggml_metal_kargs_pad_reflect_1d args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.ne02 =*/ ne02, + /*.ne03 =*/ ne03, + /*.nb00 =*/ nb00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.nb03 =*/ nb03, + /*.ne0 =*/ ne0, + /*.ne1 =*/ ne1, + /*.ne2 =*/ ne2, + /*.ne3 =*/ ne3, + /*.nb0 =*/ nb0, + /*.nb1 =*/ nb1, + /*.nb2 =*/ nb2, + /*.nb3 =*/ nb3, + /*.p0 =*/ p0, + /*.p1 =*/ p1 + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; - [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:6]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9]; - [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:11]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:12]; - [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:13]; - [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:14]; - [encoder setBytes:&p0 length:sizeof(p0) atIndex:15]; - [encoder setBytes:&p1 length:sizeof(p1) atIndex:16]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; const int nth = MIN(1024, ne0); @@ -3489,12 +3544,15 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARANGE_F32].pipeline; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_arange args = { + /*.ne0 =*/ ne0, + /*.start =*/ start, + /*.step =*/ step + }; + [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:0]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:1]; - [encoder setBytes:&start length:sizeof(start) atIndex:2]; - [encoder setBytes:&step length:sizeof(step) atIndex:3]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:0]; + [encoder setBytes:&args length:sizeof(args) atIndex:1]; const int nth = MIN(1024, ne0); @@ -3511,13 +3569,16 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32].pipeline; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_timestep_embedding args = { + /*.nb1 =*/ nb1, + /*.dim =*/ dim, + /*.max_period =*/ max_period + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:2]; - [encoder setBytes:&dim length:sizeof(dim) atIndex:3]; - [encoder setBytes:&max_period length:sizeof(max_period) atIndex:4]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; const int nth = MIN(1024, half); @@ -3550,12 +3611,15 @@ static void ggml_metal_encode_node( default: GGML_ABORT("fatal error"); }; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_argsort args = { + /*.ncols =*/ ne00, + /*.ncols_pad =*/ ne00_padded + }; + [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; - [encoder setBytes:&ne00_padded length:sizeof( int64_t) atIndex:3]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; [encoder dispatchThreadgroups:MTLSizeMake(1, nrows, 1) threadsPerThreadgroup:MTLSizeMake(ne00_padded, 1, 1)]; @@ -3569,11 +3633,14 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32].pipeline; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_leaky_relu args = { + /*.slope =*/ slope + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&slope length:sizeof(slope) atIndex:2]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; const int64_t n = ggml_nelements(dst); @@ -4149,21 +4216,24 @@ static void ggml_metal_encode_node( const int64_t n_threads = MIN((int64_t)[pipeline maxTotalThreadsPerThreadgroup], parallel_elements); const int64_t n_tg = (parallel_elements + n_threads - 1) / n_threads; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_pool_2d args_pool_2d = { + /* .k0 = */ k0, + /* .k1 = */ k1, + /* .s0 = */ s0, + /* .s1 = */ s1, + /* .p0 = */ p0, + /* .p1 = */ p1, + /* .IH = */ IH, + /* .IW = */ IW, + /* .OH = */ OH, + /* .OW = */ OW, + /* .parallel_elements = */ parallel_elements + }; + [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&k0 length:sizeof(int32_t) atIndex:2]; - [encoder setBytes:&k1 length:sizeof(int32_t) atIndex:3]; - [encoder setBytes:&s0 length:sizeof(int32_t) atIndex:4]; - [encoder setBytes:&s1 length:sizeof(int32_t) atIndex:5]; - [encoder setBytes:&p0 length:sizeof(int32_t) atIndex:6]; - [encoder setBytes:&p1 length:sizeof(int32_t) atIndex:7]; - [encoder setBytes:&IH length:sizeof(int64_t) atIndex:8]; - [encoder setBytes:&IW length:sizeof(int64_t) atIndex:9]; - [encoder setBytes:&OH length:sizeof(int64_t) atIndex:10]; - [encoder setBytes:&OW length:sizeof(int64_t) atIndex:11]; - [encoder setBytes:¶llel_elements length:sizeof(int64_t) atIndex:12]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&args_pool_2d length:sizeof(args_pool_2d) atIndex:2]; [encoder dispatchThreadgroups:MTLSizeMake(n_tg, 1, 1) threadsPerThreadgroup:MTLSizeMake(n_threads, 1, 1)]; } break; diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal index d092a16906..ad9d42a3ea 100644 --- a/ggml/src/ggml-metal/ggml-metal.metal +++ b/ggml/src/ggml-metal/ggml-metal.metal @@ -3,8 +3,7 @@ #if defined(GGML_METAL_EMBED_LIBRARY) __embed_ggml-common.h__ #else -// TODO: this should not be a relative path, but can't figure out how to set Metal include paths in Package.swift -#include "../ggml-common.h" +#include "ggml-common.h" #endif #include "ggml-metal-impl.h" @@ -948,45 +947,22 @@ kernel void kernel_cos( kernel void kernel_sum_rows( device const float * src0, device float * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant int64_t & ne03, - constant uint64_t & nb00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant uint64_t & nb03, - constant int64_t & ne10, - constant int64_t & ne11, - constant int64_t & ne12, - constant int64_t & ne13, - constant uint64_t & nb10, - constant uint64_t & nb11, - constant uint64_t & nb12, - constant uint64_t & nb13, - constant int64_t & ne0, - constant int64_t & ne1, - constant int64_t & ne2, - constant int64_t & ne3, - constant uint64_t & nb0, - constant uint64_t & nb1, - constant uint64_t & nb2, - constant uint64_t & nb3, + constant ggml_metal_kargs_sum_rows & args, uint3 tpig[[thread_position_in_grid]]) { int64_t i3 = tpig.z; int64_t i2 = tpig.y; int64_t i1 = tpig.x; - if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) { + if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) { return; } - device const float * src_row = (device const float *) ((device const char *) src0 + i1*nb01 + i2*nb02 + i3*nb03); - device float * dst_row = (device float *) ((device char *) dst + i1*nb1 + i2*nb2 + i3*nb3); + device const float * src_row = (device const float *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03); + device float * dst_row = (device float *) ((device char *) dst + i1*args.nb1 + i2*args.nb2 + i3*args.nb3); float row_sum = 0; - for (int64_t i0 = 0; i0 < ne00; i0++) { + for (int64_t i0 = 0; i0 < args.ne00; i0++) { row_sum += src_row[i0]; } @@ -998,36 +974,29 @@ kernel void kernel_soft_max( device const char * src0, device const char * src1, device char * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant float & scale, - constant float & max_bias, - constant float & m0, - constant float & m1, - constant uint32_t & n_head_log2, + constant ggml_metal_kargs_soft_max & args, threadgroup float * buf [[threadgroup(0)]], uint tgpig[[threadgroup_position_in_grid]], uint tpitg[[thread_position_in_threadgroup]], uint sgitg[[simdgroup_index_in_threadgroup]], uint tiisg[[thread_index_in_simdgroup]], uint ntg[[threads_per_threadgroup]]) { - const int64_t i03 = (tgpig) / (ne02*ne01); - const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01; - const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01); + const int64_t i03 = (tgpig) / (args.ne02*args.ne01); + const int64_t i02 = (tgpig - i03*args.ne02*args.ne01) / args.ne01; + const int64_t i01 = (tgpig - i03*args.ne02*args.ne01 - i02*args.ne01); - device const float * psrc0 = (device const float *) src0 + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); - device const T * pmask = src1 != src0 ? (device const T *) src1 + i01*ne00 : nullptr; - device float * pdst = (device float *) dst + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); + device const float * psrc0 = (device const float *) src0 + (i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00); + device const T * pmask = src1 != src0 ? (device const T *) src1 + i01*args.ne00 : nullptr; + device float * pdst = (device float *) dst + (i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00); float slope = 1.0f; // ALiBi - if (max_bias > 0.0f) { + if (args.max_bias > 0.0f) { const int64_t h = i02; - const float base = h < n_head_log2 ? m0 : m1; - const int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; + const float base = h < args.n_head_log2 ? args.m0 : args.m1; + const int exp = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1; slope = pow(base, exp); } @@ -1035,8 +1004,8 @@ kernel void kernel_soft_max( // parallel max float lmax = -INFINITY; - for (int i00 = tpitg; i00 < ne00; i00 += ntg) { - lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)); + for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) { + lmax = MAX(lmax, psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f)); } // find the max value in the block @@ -1060,8 +1029,8 @@ kernel void kernel_soft_max( // parallel sum float lsum = 0.0f; - for (int i00 = tpitg; i00 < ne00; i00 += ntg) { - const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max_val); + for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) { + const float exp_psrc0 = exp((psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f)) - max_val); lsum += exp_psrc0; pdst[i00] = exp_psrc0; } @@ -1091,7 +1060,7 @@ kernel void kernel_soft_max( const float inv_sum = 1.0f/sum; - for (int i00 = tpitg; i00 < ne00; i00 += ntg) { + for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) { pdst[i00] *= inv_sum; } } @@ -1101,35 +1070,28 @@ kernel void kernel_soft_max_4( device const char * src0, device const char * src1, device char * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant float & scale, - constant float & max_bias, - constant float & m0, - constant float & m1, - constant uint32_t & n_head_log2, + constant ggml_metal_kargs_soft_max & args, threadgroup float * buf [[threadgroup(0)]], uint tgpig[[threadgroup_position_in_grid]], uint tpitg[[thread_position_in_threadgroup]], uint sgitg[[simdgroup_index_in_threadgroup]], uint tiisg[[thread_index_in_simdgroup]], uint ntg[[threads_per_threadgroup]]) { - const int64_t i03 = (tgpig) / (ne02*ne01); - const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01; - const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01); + const int64_t i03 = (tgpig) / (args.ne02*args.ne01); + const int64_t i02 = (tgpig - i03*args.ne02*args.ne01) / args.ne01; + const int64_t i01 = (tgpig - i03*args.ne02*args.ne01 - i02*args.ne01); - device const float4 * psrc4 = (device const float4 *) src0 + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00)/4; - device const T * pmask = src1 != src0 ? (device const T *) src1 + i01*ne00/4 : nullptr; - device float4 * pdst4 = (device float4 *) dst + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00)/4; + device const float4 * psrc4 = (device const float4 *) src0 + (i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00)/4; + device const T * pmask = src1 != src0 ? (device const T *) src1 + i01*args.ne00/4 : nullptr; + device float4 * pdst4 = (device float4 *) dst + (i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00)/4; float slope = 1.0f; - if (max_bias > 0.0f) { + if (args.max_bias > 0.0f) { const int64_t h = i02; - const float base = h < n_head_log2 ? m0 : m1; - const int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; + const float base = h < args.n_head_log2 ? args.m0 : args.m1; + const int exp = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1; slope = pow(base, exp); } @@ -1137,8 +1099,8 @@ kernel void kernel_soft_max_4( // parallel max float4 lmax4 = -INFINITY; - for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) { - lmax4 = fmax(lmax4, psrc4[i00]*scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))); + for (int i00 = tpitg; i00 < args.ne00/4; i00 += ntg) { + lmax4 = fmax(lmax4, psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))); } const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3])); @@ -1163,8 +1125,8 @@ kernel void kernel_soft_max_4( // parallel sum float4 lsum4 = 0.0f; - for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) { - const float4 exp_psrc4 = exp((psrc4[i00]*scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))) - max_val); + for (int i00 = tpitg; i00 < args.ne00/4; i00 += ntg) { + const float4 exp_psrc4 = exp((psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))) - max_val); lsum4 += exp_psrc4; pdst4[i00] = exp_psrc4; } @@ -1196,7 +1158,7 @@ kernel void kernel_soft_max_4( const float inv_sum = 1.0f/sum; - for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) { + for (int i00 = tpitg; i00 < args.ne00/4; i00 += ntg) { pdst4[i00] *= inv_sum; } } @@ -1212,27 +1174,23 @@ template [[host_name("kernel_soft_max_f32_4")]] kernel kernel_soft_max_4_t kerne kernel void kernel_diag_mask_inf( device const float * src0, device float * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int & n_past, + constant ggml_metal_kargs_diag_mask_inf & args, uint3 tpig[[thread_position_in_grid]]) { const int64_t i02 = tpig[2]; const int64_t i01 = tpig[1]; const int64_t i00 = tpig[0]; - if (i00 > n_past + i01) { - dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY; + if (i00 > args.n_past + i01) { + dst[i02*args.ne01*args.ne00 + i01*args.ne00 + i00] = -INFINITY; } else { - dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00]; + dst[i02*args.ne01*args.ne00 + i01*args.ne00 + i00] = src0[i02*args.ne01*args.ne00 + i01*args.ne00 + i00]; } } kernel void kernel_diag_mask_inf_8( device const float4 * src0, device float4 * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int & n_past, + constant ggml_metal_kargs_diag_mask_inf & args, uint3 tpig[[thread_position_in_grid]]) { const int64_t i = 2*tpig[0]; @@ -1240,42 +1198,26 @@ kernel void kernel_diag_mask_inf_8( dst[i+0] = src0[i+0]; dst[i+1] = src0[i+1]; int64_t i4 = 4*i; - const int64_t i02 = i4/(ne00*ne01); i4 -= i02*ne00*ne01; - const int64_t i01 = i4/(ne00); i4 -= i01*ne00; + const int64_t i02 = i4/(args.ne00*args.ne01); i4 -= i02*args.ne00*args.ne01; + const int64_t i01 = i4/(args.ne00); i4 -= i01*args.ne00; const int64_t i00 = i4; for (int k = 3; k >= 0; --k) { - if (i00 + 4 + k <= n_past + i01) { + if (i00 + 4 + k <= args.n_past + i01) { break; } dst[i+1][k] = -INFINITY; - if (i00 + k > n_past + i01) { + if (i00 + k > args.n_past + i01) { dst[i][k] = -INFINITY; } } } // ref: ggml.c:ggml_compute_forward_ssm_conv_f32 -// TODO: optimize kernel void kernel_ssm_conv_f32( device const void * src0, device const void * src1, device float * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant uint64_t & nb00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant int64_t & ne10, - constant int64_t & ne11, - constant uint64_t & nb10, - constant uint64_t & nb11, - constant int64_t & ne0, - constant int64_t & ne1, - constant int64_t & ne2, - constant uint64_t & nb0, - constant uint64_t & nb1, - constant uint64_t & nb2, + constant ggml_metal_kargs_ssm_conv & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]], uint3 ntg[[threads_per_threadgroup]]) { @@ -1283,15 +1225,15 @@ kernel void kernel_ssm_conv_f32( const int64_t i2 = tgpig.y; const int64_t i3 = tgpig.z; - const int64_t nc = ne10; - //const int64_t ncs = ne00; - //const int64_t nr = ne01; - //const int64_t n_t = ne1; - //const int64_t n_s = ne2; + const int64_t nc = args.ne10; + //const int64_t ncs = args.ne00; + //const int64_t nr = args.ne01; + //const int64_t n_t = args.ne1; + //const int64_t n_s = args.ne2; - device const float * s = (device const float *) ((device const char *) src0 + ir*nb01 + i2*nb00 + i3*nb02); - device const float * c = (device const float *) ((device const char *) src1 + ir*nb11); - device float * x = (device float *) ((device char *) dst + ir*nb0 + i2*nb1 + i3*nb2); + device const float * s = (device const float *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02); + device const float * c = (device const float *) ((device const char *) src1 + ir*args.nb11); + device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2); float sumf = 0.0f; @@ -1303,7 +1245,6 @@ kernel void kernel_ssm_conv_f32( } // ref: ggml.c:ggml_compute_forward_ssm_scan_f32 -// TODO: optimize kernel void kernel_ssm_scan_f32( device const void * src0, device const void * src1, @@ -1312,48 +1253,27 @@ kernel void kernel_ssm_scan_f32( device const void * src4, device const void * src5, device float * dst, - constant int64_t & d_state, - constant int64_t & d_inner, - constant int64_t & n_seq_tokens, - constant int64_t & n_seqs, - constant uint64_t & nb00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant uint64_t & nb10, - constant uint64_t & nb11, - constant uint64_t & nb12, - constant uint64_t & nb13, - constant uint64_t & nb20, - constant uint64_t & nb21, - constant uint64_t & nb22, - constant uint64_t & nb30, - constant uint64_t & nb31, - constant uint64_t & nb40, - constant uint64_t & nb41, - constant uint64_t & nb42, - constant uint64_t & nb50, - constant uint64_t & nb51, - constant uint64_t & nb52, + constant ggml_metal_kargs_ssm_scan & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]], uint3 ntg[[threads_per_threadgroup]]) { const int64_t ir = tgpig.x; const int64_t i3 = tgpig.y; - const int64_t nc = d_state; - //const int64_t nr = d_inner; - const int64_t n_t = n_seq_tokens; - //const int64_t n_s = n_seqs; + const int64_t nc = args.d_state; + // const int64_t nr = args.d_inner; + const int64_t n_t = args.n_seq_tokens; + // const int64_t n_s = args.n_seqs; for (int64_t i2 = 0; i2 < n_t; ++i2) { - device const float * s0 = (device const float *) ((device const char *) src0 + ir*nb01 + i3*nb02); - device const float * x = (device const float *) ((device const char *) src1 + ir*nb10 + i2*nb11 + i3*nb12); - device const float * dt = (device const float *) ((device const char *) src2 + ir*nb20 + i2*nb21 + i3*nb22); - device const float * A = (device const float *) ((device const char *) src3 + ir*nb31); - device const float * B = (device const float *) ((device const char *) src4 + i2*nb41 + i3*nb42); - device const float * C = (device const float *) ((device const char *) src5 + i2*nb51 + i3*nb52); - device float * y = (device float *) ((device char *) dst + ir*nb10 + i2*nb11 + i3*nb12); // TODO: do not use src1 strides - device float * s = (device float *) ((device char *) dst + ir*nb01 + i3*nb02 + nb13); + device const float * s0 = (device const float *) ((device const char *) src0 + ir*args.nb01 + i3*args.nb02); + device const float * x = (device const float *) ((device const char *) src1 + ir*args.nb10 + i2*args.nb11 + i3*args.nb12); + device const float * dt = (device const float *) ((device const char *) src2 + ir*args.nb20 + i2*args.nb21 + i3*args.nb22); + device const float * A = (device const float *) ((device const char *) src3 + ir*args.nb31); + device const float * B = (device const float *) ((device const char *) src4 + i2*args.nb41 + i3*args.nb42); + device const float * C = (device const float *) ((device const char *) src5 + i2*args.nb51 + i3*args.nb52); + device float * y = (device float *) ((device char *) dst + ir*args.nb10 + i2*args.nb11 + i3*args.nb12); // TODO: do not use src1 strides + device float * s = (device float *) ((device char *) dst + ir*args.nb01 + i3*args.nb02 + args.nb13); if (i2 > 0) { s0 = s; @@ -1546,22 +1466,15 @@ kernel void kernel_rms_norm( kernel void kernel_group_norm( device const float * src0, device float * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant uint64_t & nb00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant int32_t & n_groups, - constant float & eps, + constant ggml_metal_kargs_group_norm & args, threadgroup float * buf [[threadgroup(0)]], uint tgpig[[threadgroup_position_in_grid]], uint tpitg[[thread_position_in_threadgroup]], uint sgitg[[simdgroup_index_in_threadgroup]], uint tiisg[[thread_index_in_simdgroup]], uint ntg[[threads_per_threadgroup]]) { - const int64_t ne = ne00*ne01*ne02; - const int64_t gs = ne00*ne01*((ne02 + n_groups - 1) / n_groups); + const int64_t ne = args.ne00*args.ne01*args.ne02; + const int64_t gs = args.ne00*args.ne01*((args.ne02 + args.n_groups - 1) / args.n_groups); int start = tgpig * gs; int end = start + gs; @@ -1625,7 +1538,7 @@ kernel void kernel_group_norm( } const float variance = tmp / gs; - const float scale = 1.0f/sqrt(variance + eps); + const float scale = 1.0f/sqrt(variance + args.eps); for (int j = start; j < end; j += ntg) { dst[j] *= scale; } @@ -2589,17 +2502,7 @@ template [[host_name("kernel_rope_neox_f16")]] kernel kernel_rope_neox_t kernel_ typedef void (im2col_t)( device const float * x, device char * dst, - constant int32_t & ofs0, - constant int32_t & ofs1, - constant int32_t & IW, - constant int32_t & IH, - constant int32_t & CHW, - constant int32_t & s0, - constant int32_t & s1, - constant int32_t & p0, - constant int32_t & p1, - constant int32_t & d0, - constant int32_t & d1, + constant ggml_metal_kargs_im2col & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]], uint3 tpitg[[thread_position_in_threadgroup]], @@ -2609,17 +2512,7 @@ template kernel void kernel_im2col( device const float * x, device char * dst, - constant int32_t & ofs0, - constant int32_t & ofs1, - constant int32_t & IW, - constant int32_t & IH, - constant int32_t & CHW, - constant int32_t & s0, - constant int32_t & s1, - constant int32_t & p0, - constant int32_t & p1, - constant int32_t & d0, - constant int32_t & d1, + constant ggml_metal_kargs_im2col & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]], uint3 tpitg[[thread_position_in_threadgroup]], @@ -2640,17 +2533,17 @@ kernel void kernel_im2col( const int64_t ioh = tgpig[1]; const int64_t iow = tgpig[2]; - const int64_t iiw = iow*s0 + ikw*d0 - p0; - const int64_t iih = ioh*s1 + ikh*d1 - p1; + const int64_t iiw = iow*args.s0 + ikw*args.d0 - args.p0; + const int64_t iih = ioh*args.s1 + ikh*args.d1 - args.p1; - const int64_t offset_dst = (in*OH*OW + ioh*OW + iow)*CHW + (iic*(KH*KW) + ikh*KW + ikw); + const int64_t offset_dst = (in*OH*OW + ioh*OW + iow)*args.CHW + (iic*(KH*KW) + ikh*KW + ikw); device T * pdst = (device T *) (dst); - if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) { pdst[offset_dst] = 0.0f; } else { - const int64_t offset_src = in*ofs0 + iic*ofs1 + iih*IW + iiw; + const int64_t offset_src = in*args.ofs0 + iic*args.ofs1 + iih*args.IW + iiw; pdst[offset_dst] = x[offset_src]; } } @@ -2661,20 +2554,7 @@ template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col; typedef void (im2col_ext_t)( device const float * x, device char * dst, - constant int32_t & ofs0, - constant int32_t & ofs1, - constant int32_t & IW, - constant int32_t & IH, - constant int32_t & CHW, - constant int32_t & s0, - constant int32_t & s1, - constant int32_t & p0, - constant int32_t & p1, - constant int32_t & d0, - constant int32_t & d1, - constant int32_t & N, - constant int32_t & KH, - constant int32_t & KW, + constant ggml_metal_kargs_im2col & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]], uint3 tpitg[[thread_position_in_threadgroup]], @@ -2684,53 +2564,40 @@ template kernel void kernel_im2col_ext( device const float * x, device char * dst, - constant int32_t & ofs0, - constant int32_t & ofs1, - constant int32_t & IW, - constant int32_t & IH, - constant int32_t & CHW, - constant int32_t & s0, - constant int32_t & s1, - constant int32_t & p0, - constant int32_t & p1, - constant int32_t & d0, - constant int32_t & d1, - constant int32_t & N, - constant int32_t & KH, - constant int32_t & KW, + constant ggml_metal_kargs_im2col & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]], // tgpg[0] = D x IC x KH x KW, CHW = IC x KH x KW uint3 tpitg[[thread_position_in_threadgroup]], uint3 ntg[[threads_per_threadgroup]]) { // [M, 1, 1] - const int64_t KHW = KH * KW; // KHW == ntg[1] * ntg[2], KW == ntg[2] + const int64_t KHW = (int64_t)args.KHW; - const int64_t d = tgpig[0] / CHW; - const int64_t chw = tgpig[0] % CHW; + const int64_t d = tgpig[0] / args.CHW; + const int64_t chw = tgpig[0] % args.CHW; const int64_t tgpig_0 = chw / KHW; // 0 ~ (IC - 1) const int64_t HW = tgpig[0] % KHW; const int64_t tpitg_0 = (d * ntg[0]) + tpitg[0]; - if (tpitg_0 >= N) { + if (tpitg_0 >= args.N) { return; } - const int64_t tpitg_1 = HW / KW; - const int64_t tpitg_2 = HW % KW; + const int64_t tpitg_1 = HW / args.KW; + const int64_t tpitg_2 = HW % args.KW; - const int64_t iiw = tgpig[2] * s0 + tpitg_2 * d0 - p0; - const int64_t iih = tgpig[1] * s1 + tpitg_1 * d1 - p1; + const int64_t iiw = tgpig[2] * args.s0 + tpitg_2 * args.d0 - args.p0; + const int64_t iih = tgpig[1] * args.s1 + tpitg_1 * args.d1 - args.p1; const int64_t offset_dst = - (tpitg_0 * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * CHW + - (tgpig_0 * KHW + tpitg_1 * KW + tpitg_2); + (tpitg_0 * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * args.CHW + + (tgpig_0 * KHW + tpitg_1 * args.KW + tpitg_2); device T * pdst = (device T *) (dst); - if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) { pdst[offset_dst] = 0.0f; } else { - const int64_t offset_src = tpitg_0 * ofs0 + tgpig_0 * ofs1; - pdst[offset_dst] = x[offset_src + iih * IW + iiw]; + const int64_t offset_src = tpitg_0 * args.ofs0 + tgpig_0 * args.ofs1; + pdst[offset_dst] = x[offset_src + iih * args.IW + iiw]; } } @@ -2741,12 +2608,7 @@ typedef void (conv_transpose_1d_t)( device const float * src0, device const float * src1, device char * dst, - constant int32_t & IC, - constant int32_t & IL, - constant int32_t & K, - constant int32_t & s0, - constant uint64_t & nb0, - constant uint64_t & nb1, + constant ggml_metal_kargs_conv_transpose_1d & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]]); @@ -2755,29 +2617,24 @@ kernel void kernel_conv_transpose_1d( device const T * src0, device const float * src1, device char * dst, - constant int32_t & IC, - constant int32_t & IL, - constant int32_t & K, - constant int32_t & s0, - constant uint64_t & nb0, - constant uint64_t & nb1, + constant ggml_metal_kargs_conv_transpose_1d & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]]) { float v = 0.0f; - for (int64_t c = 0; c < IC; c++) { - const int32_t kernel_offset = c * tgpg[1] * K + K * tgpig[1]; - const int32_t input_offset = c * IL; + for (int64_t c = 0; c < args.IC; c++) { + const int32_t kernel_offset = c * tgpg[1] * args.K + args.K * tgpig[1]; + const int32_t input_offset = c * args.IL; - for (int64_t i = 0; i < IL; i++) { - if (tgpig[0] >= i * s0 && tgpig[0] < i * s0 + K) { - v += src0[kernel_offset + tgpig[0] - i * s0] * src1[input_offset + i]; + for (int64_t i = 0; i < args.IL; i++) { + if (tgpig[0] >= i * args.s0 && tgpig[0] < i * args.s0 + args.K) { + v += src0[kernel_offset + tgpig[0] - i * args.s0] * src1[input_offset + i]; } } } - device float * dst_ptr = (device float *) (dst + tgpig[0] * nb0 + tgpig[1] * nb1); + device float * dst_ptr = (device float *) (dst + tgpig[0] * args.nb0 + tgpig[1] * args.nb1); dst_ptr[0] = v; } @@ -2787,12 +2644,7 @@ kernel void kernel_conv_transpose_1d( device const float * src0, device const float * src1, device char * dst, - constant int32_t & IC, - constant int32_t & IL, - constant int32_t & K, - constant int32_t & s0, - constant uint64_t & nb0, - constant uint64_t & nb1, + constant ggml_metal_kargs_conv_transpose_1d & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]]); @@ -2801,38 +2653,14 @@ kernel void kernel_conv_transpose_1d( device const half * src0, device const float * src1, device char * dst, - constant int32_t & IC, - constant int32_t & IL, - constant int32_t & K, - constant int32_t & s0, - constant uint64_t & nb0, - constant uint64_t & nb1, + constant ggml_metal_kargs_conv_transpose_1d & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]]); kernel void kernel_upscale_f32( device const char * src0, device char * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant int64_t & ne03, - constant uint64_t & nb00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant uint64_t & nb03, - constant int64_t & ne0, - constant int64_t & ne1, - constant int64_t & ne2, - constant int64_t & ne3, - constant uint64_t & nb0, - constant uint64_t & nb1, - constant uint64_t & nb2, - constant uint64_t & nb3, - constant float & sf0, - constant float & sf1, - constant float & sf2, - constant float & sf3, + constant ggml_metal_kargs_upscale & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]], uint3 ntg[[threads_per_threadgroup]]) { @@ -2841,15 +2669,15 @@ kernel void kernel_upscale_f32( const int64_t i2 = tgpig.y; const int64_t i1 = tgpig.x; - const int64_t i03 = i3/sf3; - const int64_t i02 = i2/sf2; - const int64_t i01 = i1/sf1; + const int64_t i03 = i3/args.sf3; + const int64_t i02 = i2/args.sf2; + const int64_t i01 = i1/args.sf1; - for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) { - const int64_t i00 = i0/sf0; + for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) { + const int64_t i00 = i0/args.sf0; - device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00); - device float * dst_ptr = (device float *) (dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00); + device float * dst_ptr = (device float *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0); dst_ptr[0] = src0_ptr[0]; } @@ -2858,22 +2686,7 @@ kernel void kernel_upscale_f32( kernel void kernel_pad_f32( device const char * src0, device char * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant int64_t & ne03, - constant uint64_t & nb00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant uint64_t & nb03, - constant int64_t & ne0, - constant int64_t & ne1, - constant int64_t & ne2, - constant int64_t & ne3, - constant uint64_t & nb0, - constant uint64_t & nb1, - constant uint64_t & nb2, - constant uint64_t & nb3, + constant ggml_metal_kargs_pad & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]], uint3 ntg[[threads_per_threadgroup]]) { @@ -2886,12 +2699,12 @@ kernel void kernel_pad_f32( const int64_t i02 = i2; const int64_t i01 = i1; - device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01); - device float * dst_ptr = (device float *) (dst + i3*nb3 + i2*nb2 + i1*nb1); + device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01); + device float * dst_ptr = (device float *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1); - if (i1 < ne01 && i2 < ne02 && i3 < ne03) { - for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) { - if (i0 < ne00) { + if (i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) { + for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) { + if (i0 < args.ne00) { dst_ptr[i0] = src0_ptr[i0]; } else { dst_ptr[i0] = 0.0f; @@ -2901,7 +2714,7 @@ kernel void kernel_pad_f32( return; } - for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) { + for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) { dst_ptr[i0] = 0.0f; } } @@ -2909,21 +2722,7 @@ kernel void kernel_pad_f32( kernel void kernel_pad_reflect_1d_f32( device const char * src0, device char * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant int64_t & ne03, - constant int64_t & ne0, - constant uint64_t & nb00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant uint64_t & nb03, - constant uint64_t & nb0, - constant uint64_t & nb1, - constant uint64_t & nb2, - constant uint64_t & nb3, - constant int32_t & p0, - constant int32_t & p1, + constant ggml_metal_kargs_pad_reflect_1d & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]], uint3 tpitg[[thread_position_in_threadgroup]], @@ -2937,17 +2736,17 @@ kernel void kernel_pad_reflect_1d_f32( const int64_t i02 = i2; const int64_t i01 = i1; - device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01); - device float * dst_ptr = (device float *) (dst + i3*nb3 + i2*nb2 + i1*nb1); + device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01); + device float * dst_ptr = (device float *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1); - if (i1 < ne01 && i2 < ne02 && i3 < ne03) { - for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) { - if (i0 < p0) { - dst_ptr[i0] = src0_ptr[p0 - i0]; - } else if (i0 < ne0 - p1) { - dst_ptr[i0] = src0_ptr[i0 - p0]; + if (i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) { + for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) { + if (i0 < args.p0) { + dst_ptr[i0] = src0_ptr[args.p0 - i0]; + } else if (i0 < args.ne0 - args.p1) { + dst_ptr[i0] = src0_ptr[i0 - args.p0]; } else { - dst_ptr[i0] = src0_ptr[(ne0 - p1 - p0) - (p1 + 1 - (ne0 - i0)) - 1]; + dst_ptr[i0] = src0_ptr[(args.ne0 - args.p1 - args.p0) - (args.p1 + 1 - (args.ne0 - i0)) - 1]; } } } @@ -2955,44 +2754,40 @@ kernel void kernel_pad_reflect_1d_f32( kernel void kernel_arange_f32( device char * dst, - constant int64_t & ne0, - constant float & start, - constant float & step, + constant ggml_metal_kargs_arange & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]], uint3 ntg[[threads_per_threadgroup]]) { device float * dst_ptr = (device float *) dst; - for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) { - dst_ptr[i0] = start + step * i0; + for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) { + dst_ptr[i0] = args.start + args.step * i0; } } kernel void kernel_timestep_embedding_f32( device const char * src0, device char * dst, - constant uint64_t & nb1, - constant int & dim, - constant int & max_period, + constant ggml_metal_kargs_timestep_embedding & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]], uint3 ntg[[threads_per_threadgroup]]) { int i = tgpig.x; - device float * embed_data = (device float *)(dst + i*nb1); + device float * embed_data = (device float *)(dst + i*args.nb1); - int half_ = dim / 2; + int half_ = args.dim / 2; for (int j = tpitg.x; j < half_; j += ntg.x) { float timestep = ((device float *)src0)[i]; - float freq = (float)exp(-log((float)max_period) * j / half_); + float freq = (float)exp(-log((float)args.max_period) * j / half_); float arg = timestep * freq; embed_data[j ] = cos(arg); embed_data[j + half_] = sin(arg); } - if (dim % 2 != 0 && tpitg.x == 0) { - embed_data[dim] = 0.f; + if (args.dim % 2 != 0 && tpitg.x == 0) { + embed_data[args.dim] = 0.f; } } @@ -3000,8 +2795,7 @@ kernel void kernel_timestep_embedding_f32( typedef void (argsort_t)( device const float * x, device int32_t * dst, - constant int64_t & ncols, - constant int64_t & ncols_pad, + constant ggml_metal_kargs_argsort & args, threadgroup int32_t * shared_values [[threadgroup(0)]], uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]]); @@ -3010,8 +2804,7 @@ template kernel void kernel_argsort_f32_i32( device const float * x, device int32_t * dst, - constant int64_t & ncols, - constant int64_t & ncols_pad, + constant ggml_metal_kargs_argsort & args, threadgroup int32_t * shared_values [[threadgroup(0)]], uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]]) { @@ -3019,9 +2812,9 @@ kernel void kernel_argsort_f32_i32( int col = tpitg[0]; int row = tgpig[1]; - if (col >= ncols_pad) return; + if (col >= args.ncols_pad) return; - device const float * x_row = x + row * ncols; + device const float * x_row = x + row * args.ncols; threadgroup int32_t * dst_row = shared_values; // initialize indices @@ -3029,21 +2822,21 @@ kernel void kernel_argsort_f32_i32( threadgroup_barrier(mem_flags::mem_threadgroup); - for (int k = 2; k <= ncols_pad; k *= 2) { + for (int k = 2; k <= args.ncols_pad; k *= 2) { for (int j = k / 2; j > 0; j /= 2) { int ixj = col ^ j; if (ixj > col) { if ((col & k) == 0) { - if (dst_row[col] >= ncols || - (dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ? + if (dst_row[col] >= args.ncols || + (dst_row[ixj] < args.ncols && (order == GGML_SORT_ORDER_ASC ? x_row[dst_row[col]] > x_row[dst_row[ixj]] : x_row[dst_row[col]] < x_row[dst_row[ixj]])) ) { SWAP(dst_row[col], dst_row[ixj]); } } else { - if (dst_row[ixj] >= ncols || - (dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ? + if (dst_row[ixj] >= args.ncols || + (dst_row[col] < args.ncols && (order == GGML_SORT_ORDER_ASC ? x_row[dst_row[col]] < x_row[dst_row[ixj]] : x_row[dst_row[col]] > x_row[dst_row[ixj]])) ) { @@ -3056,8 +2849,8 @@ kernel void kernel_argsort_f32_i32( } // copy the result to dst without the padding - if (col < ncols) { - dst[row * ncols + col] = dst_row[col]; + if (col < args.ncols) { + dst[row * args.ncols + col] = dst_row[col]; } } @@ -3067,9 +2860,9 @@ template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_ar kernel void kernel_leaky_relu_f32( device const float * src0, device float * dst, - constant float & slope, + constant ggml_metal_kargs_leaky_relu & args, uint tpig[[thread_position_in_grid]]) { - dst[tpig] = src0[tpig] > 0.0f ? src0[tpig] : src0[tpig] * slope; + dst[tpig] = src0[tpig] > 0.0f ? src0[tpig] : src0[tpig] * args.slope; } // ref: https://arxiv.org/pdf/2307.08691.pdf @@ -6010,28 +5803,21 @@ kernel void kernel_get_rows_q( device const void * src0, device const void * src1, device float * dst, - constant int64_t & ne00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant int64_t & ne10, - constant uint64_t & nb10, - constant uint64_t & nb11, - constant uint64_t & nb1, - constant uint64_t & nb2, + constant ggml_metal_kargs_get_rows & args, uint3 tgpig[[threadgroup_position_in_grid]], uint tiitg[[thread_index_in_threadgroup]], uint3 tptg [[threads_per_threadgroup]]) { const int64_t i10 = tgpig.x; const int64_t i11 = tgpig.y; - const int64_t r = ((const device int32_t *) ((const device char *) src1 + i11*nb11 + i10*nb10))[0]; + const int64_t r = ((const device int32_t *) ((const device char *) src1 + i11*args.nb11 + i10*args.nb10))[0]; const int64_t i02 = i11; - for (int64_t ind = tiitg; ind < ne00/16; ind += tptg.x) { + for (int64_t ind = tiitg; ind < args.ne00/16; ind += tptg.x) { float4x4 temp; - dequantize_func(((device const block_q *) ((const device char *) src0 + r*nb01 + i02*nb02)) + ind/nl, ind%nl, temp); - *(((device float4x4 *) ((device char *) dst + i11*nb2 + i10*nb1)) + ind) = temp; + dequantize_func(((device const block_q *) ((const device char *) src0 + r*args.nb01 + i02*args.nb02)) + ind/nl, ind%nl, temp); + *(((device float4x4 *) ((device char *) dst + i11*args.nb2 + i10*args.nb1)) + ind) = temp; } } @@ -6040,27 +5826,20 @@ kernel void kernel_get_rows_f( device const void * src0, device const void * src1, device float * dst, - constant int64_t & ne00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant int64_t & ne10, - constant uint64_t & nb10, - constant uint64_t & nb11, - constant uint64_t & nb1, - constant uint64_t & nb2, + constant ggml_metal_kargs_get_rows & args, uint3 tgpig[[threadgroup_position_in_grid]], uint tiitg[[thread_index_in_threadgroup]], uint3 tptg [[threads_per_threadgroup]]) { const int64_t i10 = tgpig.x; const int64_t i11 = tgpig.y; - const int64_t r = ((const device int32_t *) ((const device char *) src1 + i11*nb11 + i10*nb10))[0]; + const int64_t r = ((const device int32_t *) ((const device char *) src1 + i11*args.nb11 + i10*args.nb10))[0]; const int64_t i02 = i11; - for (int ind = tiitg; ind < ne00; ind += tptg.x) { - (( device float *) (( device char *) dst + i11*nb2 + i10*nb1))[ind] = - ((const device T *) ((const device char *) src0 + i02*nb02 + r*nb01))[ind]; + for (int ind = tiitg; ind < args.ne00; ind += tptg.x) { + (( device float *) (( device char *) dst + i11*args.nb2 + i10*args.nb1))[ind] = + ((const device T *) ((const device char *) src0 + i02*args.nb02 + r*args.nb01))[ind]; } } @@ -6068,27 +5847,20 @@ kernel void kernel_get_rows_i32( device const void * src0, device const void * src1, device int32_t * dst, - constant int64_t & ne00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant int64_t & ne10, - constant uint64_t & nb10, - constant uint64_t & nb11, - constant uint64_t & nb1, - constant uint64_t & nb2, + constant ggml_metal_kargs_get_rows & args, uint3 tgpig[[threadgroup_position_in_grid]], uint tiitg[[thread_index_in_threadgroup]], uint3 tptg [[threads_per_threadgroup]]) { const int64_t i10 = tgpig.x; const int64_t i11 = tgpig.y; - const int64_t r = ((const device int32_t *) ((const device char *) src1 + i11*nb11 + i10*nb10))[0]; + const int64_t r = ((const device int32_t *) ((const device char *) src1 + i11*args.nb11 + i10*args.nb10))[0]; const int64_t i02 = i11; - for (int ind = tiitg; ind < ne00; ind += tptg.x) { - (( device int32_t *) (( device char *) dst + i11*nb2 + i10*nb1))[ind] = - ((const device int32_t *) ((const device char *) src0 + i02*nb02 + r*nb01))[ind]; + for (int ind = tiitg; ind < args.ne00; ind += tptg.x) { + (( device int32_t *) (( device char *) dst + i11*args.nb2 + i10*args.nb1))[ind] = + ((const device int32_t *) ((const device char *) src0 + i02*args.nb02 + r*args.nb01))[ind]; } } @@ -6690,98 +6462,78 @@ template [[host_name("kernel_mul_mv_id_iq4_xs_f32")]] kernel kernel_mul_mv_id_t kernel void kernel_pool_2d_max_f32( device const float * src0, device float * dst, - constant int32_t & k0, - constant int32_t & k1, - constant int32_t & s0, - constant int32_t & s1, - constant int32_t & p0, - constant int32_t & p1, - constant int64_t & IH, - constant int64_t & IW, - constant int64_t & OH, - constant int64_t & OW, - constant int64_t & parallel_elements, + constant ggml_metal_kargs_pool_2d & args, uint gid[[thread_position_in_grid]]) { - if (gid >= parallel_elements) { + if (gid >= args.parallel_elements) { return; } const int idx = gid; - const int I_HW = IH * IW; - const int O_HW = OH * OW; + const int I_HW = args.IH * args.IW; + const int O_HW = args.OH * args.OW; const int nc = idx / O_HW; - const int cur_oh = idx % O_HW / OW; - const int cur_ow = idx % O_HW % OW; + const int cur_oh = idx % O_HW / args.OW; + const int cur_ow = idx % O_HW % args.OW; device const float * i_ptr = src0 + nc * I_HW; device float * o_ptr = dst + nc * O_HW; - const int start_h = cur_oh * s1 - p1; + const int start_h = cur_oh * args.s1 - args.p1; const int bh = MAX(0, start_h); - const int eh = MIN(IH, start_h + k1); - const int start_w = cur_ow * s0 - p0; + const int eh = MIN(args.IH, start_h + args.k1); + const int start_w = cur_ow * args.s0 - args.p0; const int bw = MAX(0, start_w); - const int ew = MIN(IW, start_w + k0); + const int ew = MIN(args.IW, start_w + args.k0); float res = -INFINITY; for (int i = bh; i < eh; i += 1) { for (int j = bw; j < ew; j += 1) { - res = MAX(res, i_ptr[i * IW + j]); + res = MAX(res, i_ptr[i * args.IW + j]); } } - o_ptr[cur_oh * OW + cur_ow] = res; + o_ptr[cur_oh * args.OW + cur_ow] = res; } kernel void kernel_pool_2d_avg_f32( device const float * src0, device float * dst, - constant int32_t & k0, - constant int32_t & k1, - constant int32_t & s0, - constant int32_t & s1, - constant int32_t & p0, - constant int32_t & p1, - constant int64_t & IH, - constant int64_t & IW, - constant int64_t & OH, - constant int64_t & OW, - constant int64_t & parallel_elements, + constant ggml_metal_kargs_pool_2d & args, uint gid[[thread_position_in_grid]]) { - if (gid >= parallel_elements) { + if (gid >= args.parallel_elements) { return; } const int idx = gid; - const int I_HW = IH * IW; - const int O_HW = OH * OW; + const int I_HW = args.IH * args.IW; + const int O_HW = args.OH * args.OW; const int nc = idx / O_HW; - const int cur_oh = idx % O_HW / OW; - const int cur_ow = idx % O_HW % OW; + const int cur_oh = idx % O_HW / args.OW; + const int cur_ow = idx % O_HW % args.OW; device const float * i_ptr = src0 + nc * I_HW; device float * o_ptr = dst + nc * O_HW; - const int start_h = cur_oh * s1 - p1; + const int start_h = cur_oh * args.s1 - args.p1; const int bh = MAX(0, start_h); - const int eh = MIN(IH, start_h + k1); - const int start_w = cur_ow * s0 - p0; + const int eh = MIN(args.IH, start_h + args.k1); + const int start_w = cur_ow * args.s0 - args.p0; const int bw = MAX(0, start_w); - const int ew = MIN(IW, start_w + k0); + const int ew = MIN(args.IW, start_w + args.k0); // const float scale = 1. / ((eh - bh) * (ew - bw)); - const float scale = 1. / (k0 * k1); + const float scale = 1. / (args.k0 * args.k1); float res = 0; for (int i = bh; i < eh; i += 1) { for (int j = bw; j < ew; j += 1) { - float cur = i_ptr[i * IW + j]; + float cur = i_ptr[i * args.IW + j]; res += cur * scale; } } - o_ptr[cur_oh * OW + cur_ow] = res; + o_ptr[cur_oh * args.OW + cur_ow] = res; } diff --git a/ggml/src/ggml-musa/CMakeLists.txt b/ggml/src/ggml-musa/CMakeLists.txt index 2c75abf61d..166970ca6b 100644 --- a/ggml/src/ggml-musa/CMakeLists.txt +++ b/ggml/src/ggml-musa/CMakeLists.txt @@ -21,7 +21,7 @@ if (MUSAToolkit_FOUND) message(STATUS "MUSA Toolkit found") if (NOT DEFINED MUSA_ARCHITECTURES) - set(MUSA_ARCHITECTURES "21;22") + set(MUSA_ARCHITECTURES "21;22;31") endif() message(STATUS "Using MUSA architectures: ${MUSA_ARCHITECTURES}") diff --git a/ggml/src/ggml-opencl/CMakeLists.txt b/ggml/src/ggml-opencl/CMakeLists.txt index 45328a6579..59a208fe9c 100644 --- a/ggml/src/ggml-opencl/CMakeLists.txt +++ b/ggml/src/ggml-opencl/CMakeLists.txt @@ -15,6 +15,7 @@ if (GGML_OPENCL_PROFILING) endif () add_compile_definitions(GGML_OPENCL_SOA_Q) +add_compile_definitions(GGML_OPENCL_TARGET_VERSION=${GGML_OPENCL_TARGET_VERSION}) if (GGML_OPENCL_USE_ADRENO_KERNELS) message(STATUS "OpenCL will use matmul kernels optimized for Adreno") diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp index dc9a718f71..14d9934fb1 100644 --- a/ggml/src/ggml-opencl/ggml-opencl.cpp +++ b/ggml/src/ggml-opencl/ggml-opencl.cpp @@ -1,4 +1,4 @@ -#define CL_TARGET_OPENCL_VERSION 220 +#define CL_TARGET_OPENCL_VERSION GGML_OPENCL_TARGET_VERSION #define CL_USE_DEPRECATED_OPENCL_1_2_APIS // suppress warnings in CL headers for GCC and Clang @@ -25,6 +25,8 @@ #include #include #include +#include +#include #undef MIN #undef MAX @@ -62,6 +64,97 @@ enum ADRENO_GPU_GEN { X1E, }; +struct ggml_cl_version { + cl_uint major = 0; + cl_uint minor = 0; +}; + +// Parses a version string of form "XX.YY ". On an error returns ggml_cl_version with all zeroes. +static ggml_cl_version parse_cl_version(std::string_view str) { + size_t major_str_begin = 0; + size_t major_str_end = str.find(".", major_str_begin); + if (major_str_end == std::string::npos) { + return {}; + } + + size_t minor_str_begin = major_str_end + 1; + size_t minor_str_end = str.find(" ", minor_str_begin); + if (minor_str_end == std::string::npos) { + return {}; + } + + cl_uint version_major; + if (std::from_chars(str.data() + major_str_begin, str.data() + major_str_end, version_major).ec != std::errc{}) { + return {}; + } + + cl_uint version_minor; + if (std::from_chars(str.data() + minor_str_begin, str.data() + minor_str_end, version_minor).ec != std::errc{}) { + return {}; + } + return { version_major, version_minor }; +} + +// Returns OpenCL platform's version. On an error returns ggml_cl_version with all zeroes. +static ggml_cl_version get_opencl_platform_version(cl_platform_id platform) { + size_t param_size; + CL_CHECK(clGetPlatformInfo(platform, CL_PLATFORM_VERSION, 0, nullptr, ¶m_size)); + std::unique_ptr param_storage(new char[param_size]); + CL_CHECK(clGetPlatformInfo(platform, CL_PLATFORM_VERSION, param_size, param_storage.get(), nullptr)); + + auto param_value = std::string_view(param_storage.get(), param_size); + const std::string version_prefix = "OpenCL "; // Suffix: "XX.YY " + if (param_value.find(version_prefix) != 0) { + return {}; + } + param_value.remove_prefix(version_prefix.length()); + return parse_cl_version(param_value); +} + +// Return a version to use in OpenCL C compilation. On an error returns ggml_cl_version with all zeroes. +static ggml_cl_version get_opencl_c_version(ggml_cl_version platform_version, cl_device_id device) { + size_t param_size; + +#if CL_TARGET_OPENCL_VERSION >= 300 + if (platform_version.major >= 3) { + CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_OPENCL_C_ALL_VERSIONS, 0, nullptr, ¶m_size)); + if (!param_size) { + return {}; + } + + std::unique_ptr versions(new cl_name_version[param_size]); + CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_OPENCL_C_ALL_VERSIONS, param_size, versions.get(), nullptr)); + unsigned versions_count = param_size / sizeof(cl_name_version); + + cl_version version_max = 0; + for (unsigned i = 0; i < versions_count; i++) { + version_max = std::max(versions[i].version, version_max); + } + + return { CL_VERSION_MAJOR(version_max), CL_VERSION_MINOR(version_max) }; + } +#else + GGML_UNUSED(platform_version); +#endif // CL_TARGET_OPENCL_VERSION >= 300 + + CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_OPENCL_C_VERSION, 0, nullptr, ¶m_size)); + if (!param_size) { + return {}; + } + + std::unique_ptr param_storage(new char[param_size]); + CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_OPENCL_C_VERSION, param_size, param_storage.get(), nullptr)); + auto param_value = std::string_view(param_storage.get(), param_size); + + const std::string version_prefix = "OpenCL C "; // Suffix: "XX.YY " + if (param_value.find(version_prefix) != 0) { + return {}; + } + param_value.remove_prefix(version_prefix.length()); + + return parse_cl_version(param_value); +} + static ADRENO_GPU_GEN get_adreno_gpu_gen(const char *device_name) { if (strstr(device_name, "730") || strstr(device_name, "740") || @@ -278,7 +371,7 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { cl_int err; -#ifdef GGML_PROFILE_OPENCL +#ifdef GGML_OPENCL_PROFILING GGML_LOG_INFO("ggml_opencl: OpenCL profiling enabled\n"); #endif @@ -470,16 +563,11 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { // A local ref of cl_device_id for convenience cl_device_id device = backend_ctx->device; - // Check device OpenCL version, OpenCL 2.0 or above is required - size_t device_ver_str_size; - clGetDeviceInfo(device, CL_DEVICE_VERSION, 0, NULL, &device_ver_str_size); - char *device_ver_buffer = (char *)alloca(device_ver_str_size + 1); - clGetDeviceInfo(device, CL_DEVICE_VERSION, device_ver_str_size, device_ver_buffer, NULL); - device_ver_buffer[device_ver_str_size] = '\0'; - GGML_LOG_INFO("ggml_opencl: device OpenCL version: %s\n", device_ver_buffer); + ggml_cl_version platform_version = get_opencl_platform_version(default_device->platform->id); - if (strstr(device_ver_buffer, "OpenCL 2") == NULL && - strstr(device_ver_buffer, "OpenCL 3") == NULL) { + // Check device OpenCL version, OpenCL 2.0 or above is required + ggml_cl_version opencl_c_version = get_opencl_c_version(platform_version, device); + if (opencl_c_version.major < 2) { GGML_LOG_ERROR("ggml_opencl: OpenCL 2.0 or above is required\n"); return backend_ctx; } @@ -516,15 +604,17 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { // If OpenCL 3.0 is supported, then check for cl_khr_subgroups, which becomes // optional in OpenCL 3.0 (cl_khr_subgroup is mandatory in OpenCL 2.x) - if (strstr(device_ver_buffer, "OpenCL 3") && - strstr(ext_buffer, "cl_khr_subgroups") == NULL && + if (opencl_c_version.major == 3 && strstr(ext_buffer, "cl_khr_subgroups") == NULL && strstr(ext_buffer, "cl_intel_subgroups") == NULL) { GGML_LOG_ERROR("ggml_opencl: device does not support subgroups (cl_khr_subgroups or cl_intel_subgroups) " "(note that subgroups is an optional feature in OpenCL 3.0)\n"); return backend_ctx; } - CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_MEM_BASE_ADDR_ALIGN, sizeof(cl_uint), &backend_ctx->alignment, NULL)); + cl_uint base_align_in_bits; + CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_MEM_BASE_ADDR_ALIGN, sizeof(cl_uint), &base_align_in_bits, NULL)); + GGML_ASSERT(base_align_in_bits % 8u == 0); + backend_ctx->alignment = base_align_in_bits / 8u; GGML_LOG_INFO("ggml_opencl: mem base addr align: %u\n", backend_ctx->alignment); clGetDeviceInfo(device, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(size_t), &backend_ctx->max_alloc_size, NULL); @@ -578,9 +668,12 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { const std::string kernel_src = read_file("ggml-opencl.cl"); #endif - std::string compile_opts = - "-cl-std=CL2.0 -cl-mad-enable -cl-unsafe-math-optimizations " - "-cl-finite-math-only -cl-fast-relaxed-math "; + auto opencl_c_std = + std::string("CL") + std::to_string(opencl_c_version.major) + "." + std::to_string(opencl_c_version.minor); + + std::string compile_opts = std::string("-cl-std=") + opencl_c_std + + " -cl-mad-enable -cl-unsafe-math-optimizations" + " -cl-finite-math-only -cl-fast-relaxed-math"; backend_ctx->program = build_program_from_source(context, device, kernel_src.c_str(), compile_opts); // Non matmul kernels. @@ -690,10 +783,10 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { CL_CHECK((backend_ctx->kernel_transpose_16 = clCreateKernel(backend_ctx->program_transpose_16, "kernel_transpose_16", &err), err)); // Gemv general - std::string CL_gemv_compile_opts = - " -cl-std=CL2.0 " - " -cl-mad-enable " - " -DSIMDGROUP_WIDTH=" + std::to_string(backend_ctx->adreno_wave_size); + std::string CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std + + " -cl-mad-enable " + " -DSIMDGROUP_WIDTH=" + + std::to_string(backend_ctx->adreno_wave_size); if (has_vector_subgroup_broadcast) { CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT "; } @@ -710,12 +803,12 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { CL_CHECK((backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_general = clCreateKernel(backend_ctx->program_CL_gemv_general, "kernel_gemv_noshuffle", &err), err)); // Gemv 2048, 16384 - CL_gemv_compile_opts = - " -cl-std=CL2.0 " - " -cl-mad-enable " - " -DLINE_STRIDE_A=2048 " - " -DBLOCK_STRIDE_A=16384 " - " -DSIMDGROUP_WIDTH=" + std::to_string(backend_ctx->adreno_wave_size); + CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std + + " -cl-mad-enable " + " -DLINE_STRIDE_A=2048 " + " -DBLOCK_STRIDE_A=16384 " + " -DSIMDGROUP_WIDTH=" + + std::to_string(backend_ctx->adreno_wave_size); if (has_vector_subgroup_broadcast) { CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT "; } @@ -732,12 +825,12 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { CL_CHECK((backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_4096_1_4096 = clCreateKernel(backend_ctx->program_CL_gemv_4096_1_4096, "kernel_gemv_noshuffle", &err), err)); // Gemv 2048, 16384 - CL_gemv_compile_opts = - " -cl-std=CL2.0 " - " -cl-mad-enable " - " -DLINE_STRIDE_A=2048 " - " -DBLOCK_STRIDE_A=16384 " - " -DSIMDGROUP_WIDTH=" + std::to_string(backend_ctx->adreno_wave_size); + CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std + + " -cl-mad-enable " + " -DLINE_STRIDE_A=2048 " + " -DBLOCK_STRIDE_A=16384 " + " -DSIMDGROUP_WIDTH=" + + std::to_string(backend_ctx->adreno_wave_size); if (has_vector_subgroup_broadcast) { CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT "; } @@ -747,12 +840,12 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { CL_CHECK((backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_4096_1_11008 = clCreateKernel(backend_ctx->program_CL_gemv_4096_1_11008, "kernel_gemv_noshuffle", &err), err)); // Gemv 5504, 44032 - CL_gemv_compile_opts = - " -cl-std=CL2.0 " - " -cl-mad-enable " - " -DLINE_STRIDE_A=5504 " - " -DBLOCK_STRIDE_A=44032 " - " -DSIMDGROUP_WIDTH=" + std::to_string(backend_ctx->adreno_wave_size); + CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std + + " -cl-mad-enable " + " -DLINE_STRIDE_A=5504 " + " -DBLOCK_STRIDE_A=44032 " + " -DSIMDGROUP_WIDTH=" + + std::to_string(backend_ctx->adreno_wave_size); if (has_vector_subgroup_broadcast) { CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT "; } @@ -762,12 +855,12 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { CL_CHECK((backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_11008_1_4096 = clCreateKernel(backend_ctx->program_CL_gemv_11008_1_4096, "kernel_gemv_noshuffle", &err), err)); // Gemv 16000, 128000 - CL_gemv_compile_opts = - " -cl-std=CL2.0 " - " -cl-mad-enable " - " -DLINE_STRIDE_A=16000 " - " -DBLOCK_STRIDE_A=128000 " - " -DSIMDGROUP_WIDTH=" + std::to_string(backend_ctx->adreno_wave_size); + CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std + + " -cl-mad-enable " + " -DLINE_STRIDE_A=16000 " + " -DBLOCK_STRIDE_A=128000 " + " -DSIMDGROUP_WIDTH=" + + std::to_string(backend_ctx->adreno_wave_size); if (has_vector_subgroup_broadcast) { CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT "; } @@ -1004,17 +1097,18 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te case GGML_OP_ADD: case GGML_OP_SCALE: case GGML_OP_MUL: - return true; + return op->src[0]->type == GGML_TYPE_F32; case GGML_OP_UNARY: switch (ggml_get_unary_op(op)) { case GGML_UNARY_OP_GELU: case GGML_UNARY_OP_SILU: case GGML_UNARY_OP_RELU: - return ggml_is_contiguous(op->src[0]); + return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32; default: return false; } case GGML_OP_CLAMP: + return op->src[0]->type == GGML_TYPE_F32; case GGML_OP_SOFT_MAX: case GGML_OP_NORM: case GGML_OP_RMS_NORM: @@ -1198,17 +1292,14 @@ struct ggml_backend_opencl_buffer_context { std::string name; }; -static void * const cl_ptr_base = (void *)(uintptr_t) 0x1000; - static void ggml_backend_opencl_buffer_free_buffer(ggml_backend_buffer_t buffer) { ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context; delete ctx; } static void * ggml_backend_opencl_buffer_get_base(ggml_backend_buffer_t buffer) { - return cl_ptr_base; - - GGML_UNUSED(buffer); + ggml_backend_opencl_context * backend_ctx = ggml_cl2_init(buffer->buft->device); + return (void *) (uintptr_t) backend_ctx->alignment; } static enum ggml_status ggml_backend_opencl_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { @@ -1241,7 +1332,7 @@ static enum ggml_status ggml_backend_opencl_buffer_init_tensor(ggml_backend_buff tensor->extra = view_extra; } else { { - size_t offset = (char *)tensor->data - (char *)cl_ptr_base; + size_t offset = (char *) tensor->data - (char *) ggml_backend_opencl_buffer_get_base(buffer); ggml_tensor_extra_cl * extra = ctx->ggml_opencl_alloc_temp_tensor_extra(); extra->offset = offset; @@ -2573,26 +2664,33 @@ static void ggml_cl_norm(ggml_backend_t backend, const ggml_tensor * src0, const memcpy(&eps, dst->op_params, sizeof(float)); const int ne00 = src0 ? src0->ne[0] : 0; - const cl_ulong nb01 = src0 ? src0->nb[1] : 0; + const int ne01 = src0 ? src0->ne[1] : 0; + const int ne02 = src0 ? src0->ne[2] : 0; + const int ne03 = src0 ? src0->ne[3] : 0; - GGML_ASSERT(ggml_is_contiguous_1(src0)); + const cl_ulong nb01 = src0 ? src0->nb[1] : 0; + const cl_ulong nb02 = src0 ? src0->nb[2] : 0; + const cl_ulong nb03 = src0 ? src0->nb[3] : 0; const int nth = MIN(64, ne00); cl_kernel kernel = backend_ctx->kernel_norm; - CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device)); - CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0)); - CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extrad->data_device)); - CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd)); - CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int), &ne00)); - CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &nb01)); - CL_CHECK(clSetKernelArg(kernel, 6, sizeof(float), &eps)); - CL_CHECK(clSetKernelArg(kernel, 7, sizeof(float)*nth, NULL)); + CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device)); + CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0)); + CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extrad->data_device)); + CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd)); + CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int), &ne00)); + CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int), &ne01)); + CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne02)); + CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int), &ne03)); + CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &nb01)); + CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb02)); + CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb03)); + CL_CHECK(clSetKernelArg(kernel, 11, sizeof(float), &eps)); + CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float)*nth, NULL)); - const int64_t nrows = ggml_nrows(src0); - - size_t global_work_size[] = {(size_t)nrows*nth, 1, 1}; + size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03}; size_t local_work_size[] = {(size_t)nth, 1, 1}; #ifdef GGML_OPENCL_PROFILING @@ -2630,16 +2728,19 @@ static void ggml_cl_rms_norm(ggml_backend_t backend, const ggml_tensor * src0, c memcpy(&eps, dst->op_params, sizeof(float)); const int ne00 = src0 ? src0->ne[0] : 0; + const int ne01 = src0 ? src0->ne[1] : 0; + const int ne02 = src0 ? src0->ne[2] : 0; + const int ne03 = src0 ? src0->ne[3] : 0; + const cl_ulong nb01 = src0 ? src0->nb[1] : 0; + const cl_ulong nb02 = src0 ? src0->nb[2] : 0; + const cl_ulong nb03 = src0 ? src0->nb[3] : 0; GGML_ASSERT(ne00 % 4 == 0); - GGML_ASSERT(ggml_is_contiguous_1(src0)); const int nth = MIN(64, ne00); - const int64_t nrows = ggml_nrows(src0); - - size_t global_work_size[] = {(size_t)nrows*nth, 1, 1}; + size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03}; size_t local_work_size[] = {(size_t)nth, 1, 1}; cl_kernel kernel = backend_ctx->kernel_rms_norm; @@ -2654,15 +2755,20 @@ static void ggml_cl_rms_norm(ggml_backend_t backend, const ggml_tensor * src0, c sizeof(local_work_size), local_work_size, sizeof(size_t), &sgs, NULL)); - CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device)); - CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0)); - CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extrad->data_device)); - CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd)); - CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int), &ne00)); - CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &nb01)); - CL_CHECK(clSetKernelArg(kernel, 6, sizeof(float), &eps)); + CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device)); + CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0)); + CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extrad->data_device)); + CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd)); + CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int), &ne00)); + CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int), &ne01)); + CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne02)); + CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int), &ne03)); + CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &nb01)); + CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb02)); + CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb03)); + CL_CHECK(clSetKernelArg(kernel, 11, sizeof(float), &eps)); // This is local memory - the size depends on subgroup size. - CL_CHECK(clSetKernelArg(kernel, 7, sizeof(float)*nth/sgs, NULL)); + CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float)*nth/sgs, NULL)); #ifdef GGML_OPENCL_PROFILING cl_event evt; @@ -3023,6 +3129,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co // enqueue kernel with profiling // <--------------------------------------------> // #ifdef GGML_OPENCL_PROFILING + cl_event evt; CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); g_profiling_info.emplace_back(); @@ -3764,10 +3871,10 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const const int ne02 = src0 ? src0->ne[2] : 0; const int ne03 = src0 ? src0->ne[3] : 0; - const int nb00 = src0 ? src0->nb[0] : 0; - const int nb01 = src0 ? src0->nb[1] : 0; - const int nb02 = src0 ? src0->nb[2] : 0; - const int nb03 = src0 ? src0->nb[3] : 0; + const cl_ulong nb00 = src0 ? src0->nb[0] : 0; + const cl_ulong nb01 = src0 ? src0->nb[1] : 0; + const cl_ulong nb02 = src0 ? src0->nb[2] : 0; + const cl_ulong nb03 = src0 ? src0->nb[3] : 0; const int ne10 = src1 ? src1->ne[0] : 0; const int ne11 = src1 ? src1->ne[1] : 0; UNUSED(ne11); @@ -3779,10 +3886,10 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const const int ne2 = dst ? dst->ne[2] : 0; const int ne3 = dst ? dst->ne[3] : 0; - const int nb0 = dst ? dst->nb[0] : 0; - const int nb1 = dst ? dst->nb[1] : 0; - const int nb2 = dst ? dst->nb[2] : 0; - const int nb3 = dst ? dst->nb[3] : 0; + const cl_ulong nb0 = dst ? dst->nb[0] : 0; + const cl_ulong nb1 = dst ? dst->nb[1] : 0; + const cl_ulong nb2 = dst ? dst->nb[2] : 0; + const cl_ulong nb3 = dst ? dst->nb[3] : 0; GGML_ASSERT(ne10 % ne02 == 0); GGML_ASSERT(ne10 >= ne02); diff --git a/ggml/src/ggml-opencl/kernels/ggml-opencl.cl b/ggml/src/ggml-opencl/kernels/ggml-opencl.cl index 8882a8c9c6..1d43642a98 100644 --- a/ggml/src/ggml-opencl/kernels/ggml-opencl.cl +++ b/ggml/src/ggml-opencl/kernels/ggml-opencl.cl @@ -506,14 +506,23 @@ kernel void kernel_norm( global float * dst, ulong offsetd, int ne00, + int ne01, + int ne02, + int ne03, ulong nb01, + ulong nb02, + ulong nb03, float eps, local float * sum ) { src0 = (global void*)((global char*)src0 + offset0); dst = (global void*)((global char*)dst + offsetd); - global float * x = (global float *) ((global char *) src0 + get_group_id(0)*nb01); + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + global float * x = (global float *) ((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01); // MEAN // parallel sum @@ -533,7 +542,7 @@ kernel void kernel_norm( // recenter and VARIANCE barrier(CLK_LOCAL_MEM_FENCE); - global float * y = dst + get_group_id(0)*ne00; + global float * y = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; sum[get_local_id(0)] = 0.0f; for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { y[i00] = x[i00] - mean; @@ -566,14 +575,23 @@ kernel void kernel_rms_norm( global float * dst, ulong offsetd, int ne00, + int ne01, + int ne02, + int ne03, ulong nb01, + ulong nb02, + ulong nb03, float eps, local float * sum // Note, the size depends on number of subgroups ) { src0 = (global void*)((global char*)src0 + offset0); dst = (global float*)((global char*)dst + offsetd); - global float4 * x = (global float4 *) ((global char *) src0 + get_group_id(0)*nb01); + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + global float4 * x = (global float4 *) ((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01); global float * x_scalar = (global float *) x; float4 sumf = 0; float all_sum = 0; @@ -607,7 +625,7 @@ kernel void kernel_rms_norm( const float mean = sum[0]; const float scale = 1.0f/sqrt(mean + eps); - global float4 * y = (global float4 *) (dst + get_group_id(0)*ne00); + global float4 * y = (global float4 *) (dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); global float * y_scalar = (global float *) y; for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) { y[i00] = x[i00] * scale; diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp index e7304947fb..6977b705e4 100644 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp @@ -41,6 +41,7 @@ #include "ggml-sycl/gemm.hpp" #include "ggml-sycl/sycl_hw.hpp" #include "ggml-sycl/getrows.hpp" +#include "ggml.h" static bool g_sycl_loaded = false; int g_ggml_sycl_debug = 0; @@ -3864,7 +3865,7 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g case GGML_UNARY_OP_GELU_QUICK: case GGML_UNARY_OP_TANH: case GGML_UNARY_OP_EXP: - return ggml_is_contiguous(op->src[0]); + return ggml_is_contiguous(op->src[0]) && (op->src[0]->type == GGML_TYPE_F32); default: return false; } @@ -3981,23 +3982,24 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g case GGML_OP_VIEW: case GGML_OP_PERMUTE: case GGML_OP_TRANSPOSE: + return true; case GGML_OP_ADD: case GGML_OP_ADD1: - case GGML_OP_LOG: case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: - return true; - case GGML_OP_NORM: - case GGML_OP_RMS_NORM: - case GGML_OP_GROUP_NORM: - return ggml_is_contiguous(op->src[0]); - case GGML_OP_SCALE: case GGML_OP_SQR: case GGML_OP_SQRT: case GGML_OP_SIN: case GGML_OP_COS: case GGML_OP_CLAMP: + case GGML_OP_LOG: + return (op->src[0]->type == GGML_TYPE_F32); + case GGML_OP_NORM: + case GGML_OP_RMS_NORM: + case GGML_OP_GROUP_NORM: + return ggml_is_contiguous(op->src[0]); + case GGML_OP_SCALE: return true; case GGML_OP_CONT: return op->src[0]->type != GGML_TYPE_BF16; diff --git a/ggml/src/ggml-sycl/mmvq.cpp b/ggml/src/ggml-sycl/mmvq.cpp index 221f65c21e..a96286d710 100644 --- a/ggml/src/ggml-sycl/mmvq.cpp +++ b/ggml/src/ggml-sycl/mmvq.cpp @@ -3,44 +3,42 @@ #include template -static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows, - const sycl::nd_item<3> &item_ct1) { - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); +static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols, const int nrows, const sycl::nd_item<3> & item_ct1) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + item_ct1.get_local_id(1); if (row >= nrows) { return; } - const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; - assert(blocks_per_warp>0); + const int blocks_per_row = ncols / qk; + constexpr int blocks_per_warp = (vdr * WARP_SIZE + qi - 1) / qi; // Ensuring blocks_per_warp > 0 -// partial sum for each thread + assert(blocks_per_warp > 0); + + // partial sum for each thread float tmp = 0.0f; - const block_q_t * x = (const block_q_t *) vx; + const block_q_t * x = (const block_q_t *) vx; const block_q8_1 * y = (const block_q8_1 *) vy; - for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; - i += blocks_per_warp) { - const int ibx = row*blocks_per_row + i; // x block index + for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; i += blocks_per_warp) { + const int ibx = row * blocks_per_row + i; // x block index - const int iby = i * (qk/QK8_1); // y block index that aligns with ibx + const int iby = i * (qk / QK8_1); // y block index that aligns with ibx - const int iqs = - vdr * - (item_ct1.get_local_id(2) % - (qi / vdr)); // x block quant index when casting the quants to int + for (size_t elem = 0; elem < qi / vdr; elem += WARP_SIZE) { + const int iqs = elem + vdr * (item_ct1.get_local_id(2) % + (qi / vdr)); // x block quant index when casting the quants to int - tmp += vec_dot_q_sycl(&x[ibx], &y[iby], iqs); + tmp += vec_dot_q_sycl(&x[ibx], &y[iby], iqs); + } } // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { + tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } if (item_ct1.get_local_id(2) == 0) { @@ -62,7 +60,7 @@ static void mul_mat_vec_q_iq2_xxs_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread @@ -87,7 +85,7 @@ static void mul_mat_vec_q_iq2_xxs_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -111,7 +109,7 @@ static void mul_mat_vec_q_iq2_xs_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread float tmp = 0.0f; @@ -135,7 +133,7 @@ static void mul_mat_vec_q_iq2_xs_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -159,7 +157,7 @@ static void mul_mat_vec_q_iq2_s_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread float tmp = 0.0f; @@ -183,7 +181,7 @@ static void mul_mat_vec_q_iq2_s_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -207,7 +205,7 @@ static void mul_mat_vec_q_iq3_xxs_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread float tmp = 0.0f; @@ -231,7 +229,7 @@ static void mul_mat_vec_q_iq3_xxs_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -255,7 +253,7 @@ static void mul_mat_vec_q_iq3_s_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread float tmp = 0.0f; @@ -279,7 +277,7 @@ static void mul_mat_vec_q_iq3_s_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -303,7 +301,7 @@ static void mul_mat_vec_q_iq1_s_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread float tmp = 0.0f; @@ -327,7 +325,7 @@ static void mul_mat_vec_q_iq1_s_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -351,7 +349,7 @@ static void mul_mat_vec_q_iq1_m_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread float tmp = 0.0f; @@ -375,7 +373,7 @@ static void mul_mat_vec_q_iq1_m_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -399,7 +397,7 @@ static void mul_mat_vec_q_iq4_nl_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread float tmp = 0.0f; @@ -423,7 +421,7 @@ static void mul_mat_vec_q_iq4_nl_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -448,7 +446,7 @@ static void mul_mat_vec_q_iq4_xs_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread float tmp = 0.0f; @@ -472,7 +470,7 @@ static void mul_mat_vec_q_iq4_xs_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -489,7 +487,7 @@ static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK4_0 == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -497,7 +495,7 @@ static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -513,7 +511,7 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK4_1 == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -521,7 +519,7 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -537,7 +535,7 @@ static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK5_0 == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -545,7 +543,7 @@ static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -561,7 +559,7 @@ static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK5_1 == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -569,7 +567,7 @@ static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -585,7 +583,7 @@ static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK8_0 == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -593,7 +591,7 @@ static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -609,7 +607,7 @@ static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -617,7 +615,7 @@ static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -633,7 +631,7 @@ static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -641,7 +639,7 @@ static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -657,7 +655,7 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -665,7 +663,7 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -681,7 +679,7 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -689,7 +687,7 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -705,7 +703,7 @@ static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -713,7 +711,7 @@ static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -730,13 +728,13 @@ static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq2_xxs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -751,13 +749,13 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler & cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq2_xs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -772,14 +770,14 @@ static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq2_s_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -794,14 +792,14 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq3_xxs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -816,14 +814,14 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq3_s_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -838,14 +836,14 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq1_s_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -860,13 +858,13 @@ static void mul_mat_vec_iq1_m_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq1_m_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -881,14 +879,14 @@ static void mul_mat_vec_iq4_nl_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK4_NL == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq4_nl_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -903,14 +901,14 @@ static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[intel::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq4_xs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp index 8cdc640e80..423ceb8a3d 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp @@ -5,23 +5,24 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; -shared FLOAT_TYPE sccache1[BLOCK_SIZE/16][16]; -shared FLOAT_TYPE sccache2[BLOCK_SIZE/16][16]; +shared FLOAT_TYPE sccache1[2][BLOCK_SIZE/16][16]; +shared FLOAT_TYPE sccache2[2][BLOCK_SIZE/16][16]; FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; +uint csel = 0; void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint v_im, const uint ix, const uint q_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) { const uint y_idx = i * QUANT_K + y_offset; [[unroll]] for (uint n = 0; n < num_rows; ++n) { const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row; + csel ^= 1; - barrier(); if (!all_threads) { // when we don't have enough blocks to use all threads if (i < num_blocks_per_row) { const uint32_t scale = uint32_t(data_a[ib0 + i].scales[itid]); - sccache1[ix][itid] = FLOAT_TYPE(scale & 0xF); - sccache2[ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF); + sccache1[csel][ix][itid] = FLOAT_TYPE(scale & 0xF); + sccache2[csel][ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF); } barrier(); @@ -29,8 +30,8 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, continue; } else { const uint32_t scale = uint32_t(data_a[ib0 + i].scales[itid]); - sccache1[ix][itid] = FLOAT_TYPE(scale & 0xF); - sccache2[ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF); + sccache1[csel][ix][itid] = FLOAT_TYPE(scale & 0xF); + sccache2[csel][ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF); barrier(); } @@ -57,22 +58,22 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, FLOAT_TYPE sum1 = FLOAT_TYPE(0.0); FLOAT_TYPE sum2 = FLOAT_TYPE(0.0); [[unroll]] for (int l = 0; l < 2; ++l) { - sum1 = fma(FLOAT_TYPE(b0[l]), sccache1[ix][ 8*v_im] * qs_u32_0[l ], - fma(FLOAT_TYPE(b16[l]), sccache1[ix][1 + 8*v_im] * qs_u32_0[l+2], - fma(FLOAT_TYPE(b32[l]), sccache1[ix][2 + 8*v_im] * qs_u32_2[l ], - fma(FLOAT_TYPE(b48[l]), sccache1[ix][3 + 8*v_im] * qs_u32_2[l+2], - fma(FLOAT_TYPE(b64[l]), sccache1[ix][4 + 8*v_im] * qs_u32_4[l ], - fma(FLOAT_TYPE(b80[l]), sccache1[ix][5 + 8*v_im] * qs_u32_4[l+2], - fma(FLOAT_TYPE(b96[l]), sccache1[ix][6 + 8*v_im] * qs_u32_6[l ], - fma(FLOAT_TYPE(b112[l]), sccache1[ix][7 + 8*v_im] * qs_u32_6[l+2], sum1)))))))); - sum2 = fma(FLOAT_TYPE(b0[l]), sccache2[ix][ 8*v_im], - fma(FLOAT_TYPE(b16[l]), sccache2[ix][1 + 8*v_im], - fma(FLOAT_TYPE(b32[l]), sccache2[ix][2 + 8*v_im], - fma(FLOAT_TYPE(b48[l]), sccache2[ix][3 + 8*v_im], - fma(FLOAT_TYPE(b64[l]), sccache2[ix][4 + 8*v_im], - fma(FLOAT_TYPE(b80[l]), sccache2[ix][5 + 8*v_im], - fma(FLOAT_TYPE(b96[l]), sccache2[ix][6 + 8*v_im], - fma(FLOAT_TYPE(b112[l]), sccache2[ix][7 + 8*v_im], sum2)))))))); + sum1 = fma(FLOAT_TYPE(b0[l]), sccache1[csel][ix][ 8*v_im] * qs_u32_0[l ], + fma(FLOAT_TYPE(b16[l]), sccache1[csel][ix][1 + 8*v_im] * qs_u32_0[l+2], + fma(FLOAT_TYPE(b32[l]), sccache1[csel][ix][2 + 8*v_im] * qs_u32_2[l ], + fma(FLOAT_TYPE(b48[l]), sccache1[csel][ix][3 + 8*v_im] * qs_u32_2[l+2], + fma(FLOAT_TYPE(b64[l]), sccache1[csel][ix][4 + 8*v_im] * qs_u32_4[l ], + fma(FLOAT_TYPE(b80[l]), sccache1[csel][ix][5 + 8*v_im] * qs_u32_4[l+2], + fma(FLOAT_TYPE(b96[l]), sccache1[csel][ix][6 + 8*v_im] * qs_u32_6[l ], + fma(FLOAT_TYPE(b112[l]), sccache1[csel][ix][7 + 8*v_im] * qs_u32_6[l+2], sum1)))))))); + sum2 = fma(FLOAT_TYPE(b0[l]), sccache2[csel][ix][ 8*v_im], + fma(FLOAT_TYPE(b16[l]), sccache2[csel][ix][1 + 8*v_im], + fma(FLOAT_TYPE(b32[l]), sccache2[csel][ix][2 + 8*v_im], + fma(FLOAT_TYPE(b48[l]), sccache2[csel][ix][3 + 8*v_im], + fma(FLOAT_TYPE(b64[l]), sccache2[csel][ix][4 + 8*v_im], + fma(FLOAT_TYPE(b80[l]), sccache2[csel][ix][5 + 8*v_im], + fma(FLOAT_TYPE(b96[l]), sccache2[csel][ix][6 + 8*v_im], + fma(FLOAT_TYPE(b112[l]), sccache2[csel][ix][7 + 8*v_im], sum2)))))))); } temp[j][n] = fma(dall, sum1, fma(-dmin, sum2, temp[j][n])); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp index 3116fad165..e91724a28d 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp @@ -5,20 +5,21 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; -shared FLOAT_TYPE sccache[BLOCK_SIZE/16][2][8]; +shared FLOAT_TYPE sccache[2][BLOCK_SIZE/16][2][8]; FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; +uint csel = 0; void calc_superblock(const uint a_offset, const uint b_offset, const uint ix, const uint itid8, const uint v_im, const uint v_im4, const uint v_in, const uint32_t hm_m[4], const uint q_offset, const uint y_offset, const uint s_shift, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) { const uint y_idx = i * QUANT_K + y_offset; [[unroll]] for (uint n = 0; n < num_rows; ++n) { const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row; + csel ^= 1; if (!all_threads) { // when we don't have enough blocks to use all threads - barrier(); if (i < num_blocks_per_row) - sccache[ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32); + sccache[csel][ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32); barrier(); if (i >= num_blocks_per_row) @@ -40,8 +41,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ix, co const vec4 qs_u32_6 = vec4(unpack8((qs_u32 >> 6) & 0x03030303)); if (all_threads) { - barrier(); - sccache[ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32); + sccache[csel][ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32); barrier(); } @@ -59,14 +59,14 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ix, co FLOAT_TYPE sum = FLOAT_TYPE(0.0); [[unroll]] for (int l = 0; l < 2; ++l) { - sum = fma(FLOAT_TYPE( b0[l]) * sccache[ix][v_im][0], qs_u32_0[l ] - hmk_0[l ], - fma(FLOAT_TYPE( b16[l]) * sccache[ix][v_im][1], qs_u32_0[l+2] - hmk_0[l+2], - fma(FLOAT_TYPE( b32[l]) * sccache[ix][v_im][2], qs_u32_2[l ] - hmk_1[l ], - fma(FLOAT_TYPE( b48[l]) * sccache[ix][v_im][3], qs_u32_2[l+2] - hmk_1[l+2], - fma(FLOAT_TYPE( b64[l]) * sccache[ix][v_im][4], qs_u32_4[l ] - hmk_2[l ], - fma(FLOAT_TYPE( b80[l]) * sccache[ix][v_im][5], qs_u32_4[l+2] - hmk_2[l+2], - fma(FLOAT_TYPE( b96[l]) * sccache[ix][v_im][6], qs_u32_6[l ] - hmk_3[l ], - fma(FLOAT_TYPE(b112[l]) * sccache[ix][v_im][7], qs_u32_6[l+2] - hmk_3[l+2], sum)))))))); + sum = fma(FLOAT_TYPE( b0[l]) * sccache[csel][ix][v_im][0], qs_u32_0[l ] - hmk_0[l ], + fma(FLOAT_TYPE( b16[l]) * sccache[csel][ix][v_im][1], qs_u32_0[l+2] - hmk_0[l+2], + fma(FLOAT_TYPE( b32[l]) * sccache[csel][ix][v_im][2], qs_u32_2[l ] - hmk_1[l ], + fma(FLOAT_TYPE( b48[l]) * sccache[csel][ix][v_im][3], qs_u32_2[l+2] - hmk_1[l+2], + fma(FLOAT_TYPE( b64[l]) * sccache[csel][ix][v_im][4], qs_u32_4[l ] - hmk_2[l ], + fma(FLOAT_TYPE( b80[l]) * sccache[csel][ix][v_im][5], qs_u32_4[l+2] - hmk_2[l+2], + fma(FLOAT_TYPE( b96[l]) * sccache[csel][ix][v_im][6], qs_u32_6[l ] - hmk_3[l ], + fma(FLOAT_TYPE(b112[l]) * sccache[csel][ix][v_im][7], qs_u32_6[l+2] - hmk_3[l+2], sum)))))))); } temp[j][n] = fma(d, sum, temp[j][n]); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp index f05f96b5ef..d53d9ee0a2 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp @@ -6,20 +6,21 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; -shared FLOAT_TYPE sccache[BLOCK_SIZE/16][16]; +shared FLOAT_TYPE sccache[2][BLOCK_SIZE/16][16]; FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; +uint csel = 0; void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint ix, const uint ql_offset, const uint qh_offset, const uint s_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) { const uint y_idx = i * QUANT_K + y_offset; [[unroll]] for (uint n = 0; n < num_rows; ++n) { const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row; + csel ^= 1; if (!all_threads) { // when we don't have enough blocks to use all threads - barrier(); if (i < num_blocks_per_row) - sccache[ix][itid] = FLOAT_TYPE(data_a[ib0 + i].scales[itid]); + sccache[csel][ix][itid] = FLOAT_TYPE(data_a[ib0 + i].scales[itid]); barrier(); if (i >= num_blocks_per_row) @@ -51,8 +52,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const vec4 q3 = vec4(unpack8(q3_u32)) - 32; if (all_threads) { - barrier(); - sccache[ix][itid] = FLOAT_TYPE(data_a[ib0 + i].scales[itid]); + sccache[csel][ix][itid] = FLOAT_TYPE(data_a[ib0 + i].scales[itid]); barrier(); } @@ -71,7 +71,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, sum[2] = fma(FLOAT_TYPE(by64[l]), q2[l], sum[2]); sum[3] = fma(FLOAT_TYPE(by96[l]), q3[l], sum[3]); } - temp[j][n] = fma(fma(sum[0], sccache[ix][s_offset], fma(sum[1], sccache[ix][s_offset + 2], fma(sum[2], sccache[ix][s_offset + 4], sum[3] * sccache[ix][s_offset + 6]))), d, temp[j][n]); + temp[j][n] = fma(fma(sum[0], sccache[csel][ix][s_offset], fma(sum[1], sccache[csel][ix][s_offset + 2], fma(sum[2], sccache[csel][ix][s_offset + 4], sum[3] * sccache[csel][ix][s_offset + 6]))), d, temp[j][n]); } } } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp index a8fd93fdea..0d03411f24 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp @@ -777,7 +777,7 @@ void main() { [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) { coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor); - [[unroll]] for (uint col = 0; col < BN; col += storestride) { + [[unroll]] for (uint col = 0; col < TN; col += storestride) { const uint row_i = dc + cm_col * TN + col + store_c; if (row_i >= _ne1) break; diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 084240331e..89409bb0e4 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -2332,6 +2332,7 @@ struct ggml_tensor * ggml_concat( struct ggml_tensor * b, int dim) { GGML_ASSERT(dim >= 0 && dim < GGML_MAX_DIMS); + GGML_ASSERT(a->type == b->type); int64_t ne[GGML_MAX_DIMS]; for (int d = 0; d < GGML_MAX_DIMS; ++d) { diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index ecac5b4bb7..19624eae04 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -253,6 +253,7 @@ class MODEL_ARCH(IntEnum): MINICPM3 = auto() GEMMA = auto() GEMMA2 = auto() + GEMMA3 = auto() STARCODER2 = auto() RWKV6 = auto() RWKV6QWEN2 = auto() @@ -440,6 +441,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.MINICPM3: "minicpm3", MODEL_ARCH.GEMMA: "gemma", MODEL_ARCH.GEMMA2: "gemma2", + MODEL_ARCH.GEMMA3: "gemma3", MODEL_ARCH.STARCODER2: "starcoder2", MODEL_ARCH.RWKV6: "rwkv6", MODEL_ARCH.RWKV6QWEN2: "rwkv6qwen2", @@ -1077,6 +1079,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_PRE_NORM, MODEL_TENSOR.FFN_POST_NORM, ], + MODEL_ARCH.GEMMA3: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_POST_NORM, + MODEL_TENSOR.FFN_PRE_NORM, + MODEL_TENSOR.FFN_POST_NORM, + ], MODEL_ARCH.STARCODER2: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, diff --git a/include/llama.h b/include/llama.h index d62792c0a6..e5286f0616 100644 --- a/include/llama.h +++ b/include/llama.h @@ -60,6 +60,7 @@ extern "C" { struct llama_model; struct llama_context; struct llama_sampler; + struct llama_kv_cache; typedef int32_t llama_pos; typedef int32_t llama_token; @@ -469,7 +470,8 @@ extern "C" { DEPRECATED(LLAMA_API int32_t llama_n_vocab (const struct llama_vocab * vocab), "use llama_vocab_n_tokens instead"); LLAMA_API const struct llama_model * llama_get_model (const struct llama_context * ctx); - LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); + LLAMA_API struct llama_kv_cache * llama_get_kv_self ( struct llama_context * ctx); + LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); // TODO: rename to llama_get_pooling_type LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model); LLAMA_API enum llama_rope_type llama_model_rope_type(const struct llama_model * model); @@ -586,7 +588,7 @@ extern "C" { // KV cache // - // TODO: remove llama_kv_cache_view_* API + // TODO: start using struct llama_kv_cache // Information associated with an individual cell in the KV cache view. struct llama_kv_cache_view_cell { @@ -641,13 +643,19 @@ extern "C" { // Returns the number of tokens in the KV cache (slow, use only for debug) // If a KV cell has multiple sequences assigned to it, it will be counted multiple times - LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx); + LLAMA_API int32_t llama_kv_self_n_tokens(const struct llama_context * ctx); + + DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx), + "use llama_kv_self_n_tokens instead"); // Returns the number of used KV cells (i.e. have at least one sequence assigned to them) - LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx); + LLAMA_API int32_t llama_kv_self_used_cells(const struct llama_context * ctx); + + DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx), + "use llama_kv_self_used_cells instead"); // Clear the KV cache - both cell info is erased and KV data is zeroed - LLAMA_API void llama_kv_cache_clear( + LLAMA_API void llama_kv_self_clear( struct llama_context * ctx); // Removes all tokens that belong to the specified sequence and have positions in [p0, p1) @@ -655,7 +663,7 @@ extern "C" { // seq_id < 0 : match any sequence // p0 < 0 : [0, p1] // p1 < 0 : [p0, inf) - LLAMA_API bool llama_kv_cache_seq_rm( + LLAMA_API bool llama_kv_self_seq_rm( struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, @@ -665,7 +673,7 @@ extern "C" { // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence // p0 < 0 : [0, p1] // p1 < 0 : [p0, inf) - LLAMA_API void llama_kv_cache_seq_cp( + LLAMA_API void llama_kv_self_seq_cp( struct llama_context * ctx, llama_seq_id seq_id_src, llama_seq_id seq_id_dst, @@ -673,17 +681,17 @@ extern "C" { llama_pos p1); // Removes all tokens that do not belong to the specified sequence - LLAMA_API void llama_kv_cache_seq_keep( + LLAMA_API void llama_kv_self_seq_keep( struct llama_context * ctx, llama_seq_id seq_id); // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) // If the KV cache is RoPEd, the KV data is updated accordingly: // - lazily on next llama_decode() - // - explicitly with llama_kv_cache_update() + // - explicitly with llama_kv_self_update() // p0 < 0 : [0, p1] // p1 < 0 : [p0, inf) - LLAMA_API void llama_kv_cache_seq_add( + LLAMA_API void llama_kv_self_seq_add( struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, @@ -693,10 +701,10 @@ extern "C" { // Integer division of the positions by factor of `d > 1` // If the KV cache is RoPEd, the KV data is updated accordingly: // - lazily on next llama_decode() - // - explicitly with llama_kv_cache_update() + // - explicitly with llama_kv_self_update() // p0 < 0 : [0, p1] // p1 < 0 : [p0, inf) - LLAMA_API void llama_kv_cache_seq_div( + LLAMA_API void llama_kv_self_seq_div( struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, @@ -704,24 +712,76 @@ extern "C" { int d); // Returns the largest position present in the KV cache for the specified sequence - LLAMA_API llama_pos llama_kv_cache_seq_pos_max( + LLAMA_API llama_pos llama_kv_self_seq_pos_max( struct llama_context * ctx, - llama_seq_id seq_id); - - // TODO: the llama_kv_cache_defrag and llama_kv_cache_update API tightly couples llama_context with llama_kv_cache - // how to avoid this? + llama_seq_id seq_id); // Defragment the KV cache // This will be applied: // - lazily on next llama_decode() - // - explicitly with llama_kv_cache_update() - LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx); - - // Apply the KV cache updates (such as K-shifts, defragmentation, etc.) - LLAMA_API void llama_kv_cache_update(struct llama_context * ctx); + // - explicitly with llama_kv_self_update() + LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx); // Check if the context supports KV cache shifting - LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx); + LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx); + + // Apply the KV cache updates (such as K-shifts, defragmentation, etc.) + LLAMA_API void llama_kv_self_update(struct llama_context * ctx); + + DEPRECATED(LLAMA_API void llama_kv_cache_clear( + struct llama_context * ctx), + "use llama_kv_self_clear instead"); + + DEPRECATED(LLAMA_API bool llama_kv_cache_seq_rm( + struct llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1), + "use llama_kv_self_seq_rm instead"); + + DEPRECATED(LLAMA_API void llama_kv_cache_seq_cp( + struct llama_context * ctx, + llama_seq_id seq_id_src, + llama_seq_id seq_id_dst, + llama_pos p0, + llama_pos p1), + "use llama_kv_self_seq_cp instead"); + + DEPRECATED(LLAMA_API void llama_kv_cache_seq_keep( + struct llama_context * ctx, + llama_seq_id seq_id), + "use llama_kv_self_seq_keep instead"); + + DEPRECATED(LLAMA_API void llama_kv_cache_seq_add( + struct llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + llama_pos delta), + "use llama_kv_self_seq_add instead"); + + DEPRECATED(LLAMA_API void llama_kv_cache_seq_div( + struct llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + int d), + "use llama_kv_self_seq_div instead"); + + DEPRECATED(LLAMA_API llama_pos llama_kv_cache_seq_pos_max( + struct llama_context * ctx, + llama_seq_id seq_id), + "use llama_kv_self_seq_pos_max instead"); + + DEPRECATED(LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx), + "use llama_kv_self_defrag instead"); + + DEPRECATED(LLAMA_API bool llama_kv_cache_can_shift(const struct llama_context * ctx), + "use llama_kv_self_can_shift instead"); + + DEPRECATED(LLAMA_API void llama_kv_cache_update(struct llama_context * ctx), + "use llama_kv_self_update instead"); + // // State / sessions diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last index 040b53ca39..c7944d1d42 100644 --- a/scripts/sync-ggml.last +++ b/scripts/sync-ggml.last @@ -1 +1 @@ -58ecf6b96d887e408b6869915863fa1126483d51 +c7dfe3d174f98b14801f9ed12f129179d3e7b638 diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt index e1b02e4c08..b340dae5b2 100644 --- a/src/CMakeLists.txt +++ b/src/CMakeLists.txt @@ -15,18 +15,21 @@ add_library(llama llama-chat.cpp llama-context.cpp llama-grammar.cpp + llama-graph.cpp llama-hparams.cpp llama-impl.cpp + llama-io.cpp llama-kv-cache.cpp + llama-memory.cpp llama-mmap.cpp llama-model-loader.cpp llama-model.cpp llama-quant.cpp llama-sampling.cpp llama-vocab.cpp - unicode.h - unicode.cpp unicode-data.cpp + unicode.cpp + unicode.h ) target_include_directories(llama PUBLIC . ../include ../common) diff --git a/src/llama-adapter.cpp b/src/llama-adapter.cpp index 8a08004631..b448614e47 100644 --- a/src/llama-adapter.cpp +++ b/src/llama-adapter.cpp @@ -4,14 +4,13 @@ #include "llama-mmap.h" #include "llama-model.h" -#include #include #include #include // vec -struct ggml_tensor * llama_adapter_cvec::tensor_for(int il) const { +ggml_tensor * llama_adapter_cvec::tensor_for(int il) const { if (il < 0 || il < layer_start || il > layer_end || (size_t) il >= tensors.size()) { return nullptr; } @@ -19,7 +18,7 @@ struct ggml_tensor * llama_adapter_cvec::tensor_for(int il) const { return tensors[il]; } -struct ggml_tensor * llama_adapter_cvec::apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int il) const { +ggml_tensor * llama_adapter_cvec::apply_to(ggml_context * ctx, ggml_tensor * cur, int il) const { ggml_tensor * layer_dir = tensor_for(il); if (layer_dir != nullptr) { cur = ggml_add(ctx, cur, layer_dir); @@ -40,7 +39,7 @@ bool llama_adapter_cvec::init(const llama_model & model) { auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * { auto it = ctx_map.find(buft); if (it == ctx_map.end()) { - struct ggml_init_params params = { + ggml_init_params params = { /*.mem_size =*/ hparams.n_layer*ggml_tensor_overhead(), /*.mem_buffer =*/ NULL, /*.no_alloc =*/ true, @@ -91,7 +90,7 @@ bool llama_adapter_cvec::init(const llama_model & model) { return true; } -int32_t llama_adapter_cvec::apply( +bool llama_adapter_cvec::apply( const llama_model & model, const float * data, size_t len, @@ -104,17 +103,17 @@ int32_t llama_adapter_cvec::apply( // disable the current control vector (but leave allocated for later) layer_start = -1; layer_end = -1; - return 0; + return true; } if (n_embd != (int) hparams.n_embd) { LLAMA_LOG_ERROR("%s: control vector n_embd does not match model\n", __func__); - return 1; + return false; } if (tensors.empty()) { if (!init(model)) { - return 1; + return false; } } @@ -130,12 +129,12 @@ int32_t llama_adapter_cvec::apply( } } - return 0; + return true; } // lora -llama_adapter_lora_weight * llama_adapter_lora::get_weight(struct ggml_tensor * w) { +llama_adapter_lora_weight * llama_adapter_lora::get_weight(ggml_tensor * w) { const std::string name(w->name); const auto pos = ab_map.find(name); @@ -146,11 +145,11 @@ llama_adapter_lora_weight * llama_adapter_lora::get_weight(struct ggml_tensor * return nullptr; } -static void llama_adapter_lora_init_impl(struct llama_model & model, const char * path_lora, struct llama_adapter_lora & adapter) { +static void llama_adapter_lora_init_impl(llama_model & model, const char * path_lora, llama_adapter_lora & adapter) { LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora); ggml_context * ctx_init; - struct gguf_init_params meta_gguf_params = { + gguf_init_params meta_gguf_params = { /* .no_alloc = */ true, /* .ctx = */ &ctx_init, }; @@ -201,7 +200,7 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char auto it = ctx_map.find(buft); if (it == ctx_map.end()) { // add a new context - struct ggml_init_params params = { + ggml_init_params params = { /*.mem_size =*/ n_tensors*ggml_tensor_overhead(), /*.mem_buffer =*/ NULL, /*.no_alloc =*/ true, @@ -264,7 +263,7 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model (hint: maybe wrong base model?)"); } - struct ggml_context * dev_ctx = ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer)); + ggml_context * dev_ctx = ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer)); // validate tensor shape if (is_token_embd) { // expect B to be non-transposed, A and B are flipped; see llm_build_inp_embd() @@ -281,8 +280,8 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char } // save tensor to adapter - struct ggml_tensor * tensor_a = ggml_dup_tensor(dev_ctx, w.a); - struct ggml_tensor * tensor_b = ggml_dup_tensor(dev_ctx, w.b); + ggml_tensor * tensor_a = ggml_dup_tensor(dev_ctx, w.a); + ggml_tensor * tensor_b = ggml_dup_tensor(dev_ctx, w.b); ggml_set_name(tensor_a, w.a->name); ggml_set_name(tensor_b, w.b->name); adapter.ab_map[name] = llama_adapter_lora_weight(tensor_a, tensor_b); @@ -308,7 +307,7 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char { llama_file gguf_file(path_lora, "rb"); std::vector read_buf; - auto set_tensor = [&](struct ggml_tensor * orig, struct ggml_tensor * dev) { + auto set_tensor = [&](ggml_tensor * orig, ggml_tensor * dev) { size_t offs = gguf_get_data_offset(ctx_gguf.get()) + gguf_get_tensor_offset(ctx_gguf.get(), gguf_find_tensor(ctx_gguf.get(), orig->name)); size_t size = ggml_nbytes(orig); read_buf.resize(size); @@ -327,8 +326,8 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2); } -struct llama_adapter_lora * llama_adapter_lora_init(struct llama_model * model, const char * path_lora) { - struct llama_adapter_lora * adapter = new llama_adapter_lora(); +llama_adapter_lora * llama_adapter_lora_init(llama_model * model, const char * path_lora) { + llama_adapter_lora * adapter = new llama_adapter_lora(); try { llama_adapter_lora_init_impl(*model, path_lora, *adapter); @@ -342,6 +341,6 @@ struct llama_adapter_lora * llama_adapter_lora_init(struct llama_model * model, return nullptr; } -void llama_adapter_lora_free(struct llama_adapter_lora * adapter) { +void llama_adapter_lora_free(llama_adapter_lora * adapter) { delete adapter; } diff --git a/src/llama-adapter.h b/src/llama-adapter.h index 603fa08f6d..65824e9727 100644 --- a/src/llama-adapter.h +++ b/src/llama-adapter.h @@ -15,11 +15,11 @@ // struct llama_adapter_cvec { - struct ggml_tensor * tensor_for(int il) const; + ggml_tensor * tensor_for(int il) const; - struct ggml_tensor * apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int il) const; + ggml_tensor * apply_to(ggml_context * ctx, ggml_tensor * cur, int il) const; - int32_t apply( + bool apply( const llama_model & model, const float * data, size_t len, @@ -36,7 +36,7 @@ private: std::vector ctxs; std::vector bufs; - std::vector tensors; // per layer + std::vector tensors; // per layer }; // @@ -44,8 +44,8 @@ private: // struct llama_adapter_lora_weight { - struct ggml_tensor * a = nullptr; - struct ggml_tensor * b = nullptr; + ggml_tensor * a = nullptr; + ggml_tensor * b = nullptr; // get actual scale based on rank and alpha float get_scale(float alpha, float adapter_scale) const { @@ -55,12 +55,12 @@ struct llama_adapter_lora_weight { } llama_adapter_lora_weight() = default; - llama_adapter_lora_weight(struct ggml_tensor * a, struct ggml_tensor * b) : a(a), b(b) {} + llama_adapter_lora_weight(ggml_tensor * a, ggml_tensor * b) : a(a), b(b) {} }; struct llama_adapter_lora { // map tensor name to lora_a_b - std::unordered_map ab_map; + std::unordered_map ab_map; std::vector ctxs; std::vector bufs; @@ -70,5 +70,7 @@ struct llama_adapter_lora { llama_adapter_lora() = default; ~llama_adapter_lora() = default; - llama_adapter_lora_weight * get_weight(struct ggml_tensor * w); + llama_adapter_lora_weight * get_weight(ggml_tensor * w); }; + +using llama_adapter_loras = std::unordered_map; diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index 97a1e7e5e0..28f2bbc8f7 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -36,6 +36,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_MINICPM3, "minicpm3" }, { LLM_ARCH_GEMMA, "gemma" }, { LLM_ARCH_GEMMA2, "gemma2" }, + { LLM_ARCH_GEMMA3, "gemma3" }, { LLM_ARCH_STARCODER2, "starcoder2" }, { LLM_ARCH_MAMBA, "mamba" }, { LLM_ARCH_XVERSE, "xverse" }, @@ -766,6 +767,26 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" }, }, }, + { + LLM_ARCH_GEMMA3, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" }, + }, + }, { LLM_ARCH_STARCODER2, { diff --git a/src/llama-arch.h b/src/llama-arch.h index 122fdcebe0..2ec2e2362e 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -40,6 +40,7 @@ enum llm_arch { LLM_ARCH_MINICPM3, LLM_ARCH_GEMMA, LLM_ARCH_GEMMA2, + LLM_ARCH_GEMMA3, LLM_ARCH_STARCODER2, LLM_ARCH_MAMBA, LLM_ARCH_XVERSE, diff --git a/src/llama-batch.h b/src/llama-batch.h index 773c3808b7..f1df40d270 100644 --- a/src/llama-batch.h +++ b/src/llama-batch.h @@ -42,9 +42,9 @@ struct llama_sbatch { bool logits_all; // TODO: remove once lctx.logits_all is removed too // sorted indices into the batch - std::vector ids; + std::vector ids; // batch indices of the output - std::vector out_ids; + std::vector out_ids; std::vector seq; const llama_batch * batch = nullptr; diff --git a/src/llama-context.cpp b/src/llama-context.cpp index 671d2a81ad..89fb33cbcd 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -1,551 +1,1566 @@ #include "llama-context.h" #include "llama-impl.h" +#include "llama-io.h" #include "llama-mmap.h" +#include "llama-model.h" +#include "llama-kv-cache.h" #include -#include #include #include +#include -void llama_set_k_shift(struct llama_context & lctx) { - const int64_t kv_size = lctx.kv_self.size; +// +// llama_context +// - assert(ggml_backend_buffer_is_host(lctx.inp_K_shift->buffer)); +llama_context::llama_context( + const llama_model & model, + llama_context_params params) : + model(model) { + LLAMA_LOG_INFO("%s: constructing llama_context\n", __func__); - int32_t * data = (int32_t *) lctx.inp_K_shift->data; + t_start_us = model.t_start_us; + t_load_us = model.t_load_us; - for (int i = 0; i < kv_size; ++i) { - data[i] = lctx.kv_self.cells[i].delta; - } -} + const auto & hparams = model.hparams; -void llama_set_s_copy(struct llama_context & lctx) { - const int64_t kv_size = lctx.kv_self.size; + cparams.n_seq_max = std::max(1u, params.n_seq_max); + cparams.n_threads = params.n_threads; + cparams.n_threads_batch = params.n_threads_batch; + cparams.yarn_ext_factor = params.yarn_ext_factor; + cparams.yarn_attn_factor = params.yarn_attn_factor; + cparams.yarn_beta_fast = params.yarn_beta_fast; + cparams.yarn_beta_slow = params.yarn_beta_slow; + cparams.defrag_thold = params.defrag_thold; + cparams.embeddings = params.embeddings; + cparams.offload_kqv = params.offload_kqv; + cparams.flash_attn = params.flash_attn; + cparams.no_perf = params.no_perf; + cparams.pooling_type = params.pooling_type; - assert(ggml_backend_buffer_is_host(lctx.inp_s_copy->buffer)); + cparams.n_ctx = params.n_ctx == 0 ? hparams.n_ctx_train : params.n_ctx; + cparams.rope_freq_base = params.rope_freq_base == 0.0f ? hparams.rope_freq_base_train : params.rope_freq_base; + cparams.rope_freq_scale = params.rope_freq_scale == 0.0f ? hparams.rope_freq_scale_train : params.rope_freq_scale; - int32_t * data = (int32_t *) lctx.inp_s_copy->data; + cparams.n_ctx_orig_yarn = params.yarn_orig_ctx != 0 ? params.yarn_orig_ctx : + hparams.n_ctx_orig_yarn != 0 ? hparams.n_ctx_orig_yarn : + hparams.n_ctx_train; - for (int i = 0; i < kv_size; ++i) { - data[i] = lctx.kv_self.cells[i].src; - } -} + cparams.cb_eval = params.cb_eval; + cparams.cb_eval_user_data = params.cb_eval_user_data; -// llama input - -static int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buckets, bool bidirectional) { - // TODO move to hparams if a T5 variant appears that uses a different value - const int64_t max_distance = 128; - - if (bidirectional) { - n_buckets >>= 1; + auto rope_scaling_type = params.rope_scaling_type; + if (rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED) { + rope_scaling_type = hparams.rope_scaling_type_train; } - const int64_t max_exact = n_buckets >> 1; + if (rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_NONE) { + cparams.rope_freq_scale = 1.0f; // never scale if scaling type is none + } - int32_t relative_position = x - y; - int32_t relative_bucket = 0; - if (bidirectional) { - relative_bucket += (relative_position > 0) * n_buckets; - relative_position = abs(relative_position); + if (cparams.yarn_ext_factor < 0.0f) { // negative indicates 'not set' + cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_YARN ? 1.0f : 0.0f; + } + + cparams.yarn_attn_factor *= hparams.rope_attn_factor; + + if (cparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) { + if (hparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) { + cparams.pooling_type = LLAMA_POOLING_TYPE_NONE; + } else { + cparams.pooling_type = hparams.pooling_type; + } + } + + if (params.attention_type == LLAMA_ATTENTION_TYPE_UNSPECIFIED) { + cparams.causal_attn = hparams.causal_attn; } else { - relative_position = -std::min(relative_position, 0); + cparams.causal_attn = params.attention_type == LLAMA_ATTENTION_TYPE_CAUSAL; + } + + // with causal attention, the batch size is limited by the context size + cparams.n_batch = cparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch; + + // the batch has to be at least GGML_KQ_MASK_PAD because we will be padding the KQ_mask + // this is required by GPU kernels in order to avoid out-of-bounds accesses (e.g. ggml_flash_attn_ext) + // ref: https://github.com/ggerganov/llama.cpp/pull/5021 + // TODO: this padding is not needed for the cache-less context so we should probably move it to llama_context_kv_self + if (cparams.n_batch < GGML_KQ_MASK_PAD) { + LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD); + cparams.n_batch = GGML_KQ_MASK_PAD; + } + + cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch); + + const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max; + + LLAMA_LOG_INFO("%s: n_seq_max = %u\n", __func__, cparams.n_seq_max); + LLAMA_LOG_INFO("%s: n_ctx = %u\n", __func__, cparams.n_ctx); + LLAMA_LOG_INFO("%s: n_ctx_per_seq = %u\n", __func__, n_ctx_per_seq); + LLAMA_LOG_INFO("%s: n_batch = %u\n", __func__, cparams.n_batch); + LLAMA_LOG_INFO("%s: n_ubatch = %u\n", __func__, cparams.n_ubatch); + LLAMA_LOG_INFO("%s: causal_attn = %d\n", __func__, cparams.causal_attn); + LLAMA_LOG_INFO("%s: flash_attn = %d\n", __func__, cparams.flash_attn); + LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, cparams.rope_freq_base); + LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, cparams.rope_freq_scale); + + if (n_ctx_per_seq < hparams.n_ctx_train) { + LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n", + __func__, n_ctx_per_seq, hparams.n_ctx_train); + } + + if (n_ctx_per_seq > hparams.n_ctx_train) { + LLAMA_LOG_WARN("%s: n_ctx_pre_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n", + __func__, n_ctx_per_seq, hparams.n_ctx_train); + } + + logits_all = params.logits_all; + + if (!hparams.vocab_only) { + // GPU backends + for (auto * dev : model.devices) { + ggml_backend_t backend = ggml_backend_dev_init(dev, nullptr); + if (backend == nullptr) { + throw std::runtime_error(format("failed to initialize %s backend", ggml_backend_dev_name(dev))); + } + backends.emplace_back(backend); + } + + // add ACCEL backends (such as BLAS) + for (size_t i = 0; i < ggml_backend_dev_count(); ++i) { + ggml_backend_dev_t dev = ggml_backend_dev_get(i); + if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_ACCEL) { + ggml_backend_t backend = ggml_backend_dev_init(dev, nullptr); + if (backend == nullptr) { + throw std::runtime_error(format("failed to initialize %s backend", ggml_backend_dev_name(dev))); + } + backends.emplace_back(backend); + } + } + + // add CPU backend + backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr); + if (backend_cpu == nullptr) { + throw std::runtime_error("failed to initialize CPU backend"); + } + backends.emplace_back(backend_cpu); + + // create a list of the set_n_threads functions in the backends + for (auto & backend : backends) { + ggml_backend_dev_t dev = ggml_backend_get_device(backend.get()); + ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr; + if (reg) { + auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads"); + if (ggml_backend_set_n_threads_fn) { + set_n_threads_fns.emplace_back(backend.get(), ggml_backend_set_n_threads_fn); + } + } + } + + llama_set_abort_callback(this, params.abort_callback, params.abort_callback_data); + + // graph outputs buffer + { + // resized during inference when a batch uses more outputs + if ((uint32_t) output_reserve(params.n_seq_max) < params.n_seq_max) { + throw std::runtime_error("failed to reserve initial output buffer"); + } + + LLAMA_LOG_INFO("%s: %10s output buffer size = %8.2f MiB\n", __func__, + ggml_backend_buffer_name (buf_output.get()), + ggml_backend_buffer_get_size(buf_output.get()) / 1024.0 / 1024.0); + } + } + + // init the memory module + // TODO: for now, always create a unified KV cache + if (!hparams.vocab_only) { + kv_self.reset(static_cast(model.create_memory())); + + LLAMA_LOG_DEBUG("%s: n_ctx = %u\n", __func__, cparams.n_ctx); + + cparams.n_ctx = GGML_PAD(cparams.n_ctx, kv_self->get_padding(cparams)); + + LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx); + + uint32_t kv_size = cparams.n_ctx; + ggml_type type_k = params.type_k; + ggml_type type_v = params.type_v; + + if (llama_model_is_recurrent(&model)) { + // Mamba needs at least as many KV cells as there are sequences kept at any time + kv_size = std::max((uint32_t) 1, params.n_seq_max); + // it's probably best to keep as much precision as possible for the states + type_k = GGML_TYPE_F32; // required by ggml_ssm_conv for Mamba's conv_states + type_v = GGML_TYPE_F32; // required by ggml_ssm_scan for Mamba's ssm_states + } + + GGML_ASSERT(hparams.n_embd_head_k % ggml_blck_size(type_k) == 0); + GGML_ASSERT(hparams.n_embd_head_v % ggml_blck_size(type_v) == 0); + + if (!kv_self->init(model, cparams, type_k, type_v, kv_size, cparams.offload_kqv)) { + throw std::runtime_error("failed to initialize self-attention cache"); + } + + { + const size_t memory_size_k = kv_self->size_k_bytes(); + const size_t memory_size_v = kv_self->size_v_bytes(); + + LLAMA_LOG_INFO("%s: KV self size = %7.2f MiB, K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__, + (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f), + ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f), + ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f)); + } + } + + // init backends + if (!hparams.vocab_only) { + LLAMA_LOG_DEBUG("%s: enumerating backends\n", __func__); + + backend_buft.clear(); + backend_ptrs.clear(); + + for (auto & backend : backends) { + auto * buft = ggml_backend_get_default_buffer_type(backend.get()); + auto backend_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get())); + + if (backend_type == GGML_BACKEND_DEVICE_TYPE_CPU && !model.devices.empty()) { + // use the host buffer of the first device CPU for faster transfer of the intermediate state + auto * dev = model.devices[0]; + auto * host_buft = ggml_backend_dev_host_buffer_type(dev); + if (host_buft) { + buft = host_buft; + } + } + + backend_buft.push_back(buft); + backend_ptrs.push_back(backend.get()); + } + + LLAMA_LOG_DEBUG("%s: backend_ptrs.size() = %zu\n", __func__, backend_ptrs.size()); + + const size_t max_nodes = this->graph_max_nodes(); + + LLAMA_LOG_DEBUG("%s: max_nodes = %zu\n", __func__, max_nodes); + + // buffer used to store the computation graph and the tensor meta data + buf_compute_meta.resize(ggml_tensor_overhead()*max_nodes + ggml_graph_overhead_custom(max_nodes, false)); + + // TODO: move these checks to ggml_backend_sched + // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary + bool pipeline_parallel = + model.n_devices() > 1 && + model.params.n_gpu_layers > (int) model.hparams.n_layer && + model.params.split_mode == LLAMA_SPLIT_MODE_LAYER && + cparams.offload_kqv; + + // pipeline parallelism requires support for async compute and events in all devices + if (pipeline_parallel) { + for (auto & backend : backends) { + auto dev_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get())); + if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU) { + // ignore CPU backend + continue; + } + auto * dev = ggml_backend_get_device(backend.get()); + ggml_backend_dev_props props; + ggml_backend_dev_get_props(dev, &props); + if (!props.caps.async || !props.caps.events) { + // device does not support async compute or events + pipeline_parallel = false; + break; + } + } + } + + sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, pipeline_parallel)); + + if (pipeline_parallel) { + LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(sched.get())); + } + } + + // reserve worst-case graph + if (!hparams.vocab_only) { + uint32_t n_seqs = 1; // TODO: worst-case number of sequences + uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch); + + llama_token token = model.vocab.token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph + + // max number of outputs + n_outputs = n_tokens; + + LLAMA_LOG_DEBUG("%s: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs); + + int n_splits_pp = -1; + int n_nodes_pp = -1; + + int n_splits_tg = -1; + int n_nodes_tg = -1; + + // simulate full KV cache + kv_self->n = kv_self->size; + + cross.v_embd.clear(); + + // reserve pp graph first so that buffers are only allocated once + { + llama_ubatch ubatch_pp = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr}; + auto * gf = graph_init(); + graph_build(ctx_compute.get(), gf, ubatch_pp, LLM_GRAPH_TYPE_DEFAULT); + if (!ggml_backend_sched_reserve(sched.get(), gf)) { + throw std::runtime_error("failed to allocate compute pp buffers"); + } + + n_splits_pp = ggml_backend_sched_get_n_splits(sched.get()); + n_nodes_pp = ggml_graph_n_nodes(gf); + } + + // reserve with tg graph to get the number of splits and nodes + { + llama_ubatch ubatch_tg = { true, 1, 1, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr}; + auto * gf = graph_init(); + graph_build(ctx_compute.get(), gf, ubatch_tg, LLM_GRAPH_TYPE_DEFAULT); + if (!ggml_backend_sched_reserve(sched.get(), gf)) { + throw std::runtime_error("failed to allocate compute tg buffers"); + } + n_splits_tg = ggml_backend_sched_get_n_splits(sched.get()); + n_nodes_tg = ggml_graph_n_nodes(gf); + } + + // reserve again with pp graph to avoid ggml-alloc reallocations during inference + { + llama_ubatch ubatch_pp = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr}; + auto * gf = graph_init(); + graph_build(ctx_compute.get(), gf, ubatch_pp, LLM_GRAPH_TYPE_DEFAULT); + if (!ggml_backend_sched_reserve(sched.get(), gf)) { + throw std::runtime_error("failed to allocate compute pp buffers"); + } + } + + for (size_t i = 0; i < backend_ptrs.size(); ++i) { + ggml_backend_t backend = backend_ptrs[i]; + ggml_backend_buffer_type_t buft = backend_buft[i]; + size_t size = ggml_backend_sched_get_buffer_size(sched.get(), backend); + if (size > 1) { + LLAMA_LOG_INFO("%s: %10s compute buffer size = %8.2f MiB\n", __func__, + ggml_backend_buft_name(buft), + size / 1024.0 / 1024.0); + } + } + + if (n_nodes_pp == n_nodes_tg) { + LLAMA_LOG_INFO("%s: graph nodes = %d\n", __func__, n_nodes_pp); + } else { + LLAMA_LOG_INFO("%s: graph nodes = %d (with bs=%d), %d (with bs=1)\n", __func__, n_nodes_pp, n_tokens, n_nodes_tg); + } + + if (n_splits_pp == n_splits_tg) { + LLAMA_LOG_INFO("%s: graph splits = %d\n", __func__, n_splits_pp); + } else { + LLAMA_LOG_INFO("%s: graph splits = %d (with bs=%d), %d (with bs=1)\n", __func__, n_splits_pp, n_tokens, n_splits_tg); + } } - int32_t relative_position_if_large = floorf(max_exact + logf(1.0 * relative_position / max_exact) * (n_buckets - max_exact) / log(1.0 * max_distance / max_exact)); - relative_position_if_large = std::min(relative_position_if_large, n_buckets - 1); - relative_bucket += (relative_position < max_exact ? relative_position : relative_position_if_large); - return relative_bucket; } -void llama_set_inputs(llama_context & lctx, const llama_ubatch & ubatch) { +llama_context::~llama_context() = default; + +void llama_context::synchronize() { + ggml_backend_sched_synchronize(sched.get()); + + // FIXME: if multiple single tokens are evaluated without a synchronization, + // the stats will be added to the prompt evaluation stats + // this should only happen when using batch size 1 to evaluate a batch + + // add the evaluation to the stats + if (n_queued_tokens == 1) { + if (!cparams.no_perf) { + t_eval_us += ggml_time_us() - t_compute_start_us; + } + n_eval++; + } else if (n_queued_tokens > 1) { + if (!cparams.no_perf) { + t_p_eval_us += ggml_time_us() - t_compute_start_us; + } + n_p_eval += n_queued_tokens; + } + + // get a more accurate load time, upon first eval + if (n_queued_tokens > 0 && !has_evaluated_once) { + t_load_us = ggml_time_us() - t_start_us; + has_evaluated_once = true; + } + + n_queued_tokens = 0; + t_compute_start_us = 0; +} + +const llama_model & llama_context::get_model() const { + return model; +} + +uint32_t llama_context::n_ctx() const { + return cparams.n_ctx; +} + +uint32_t llama_context::n_ctx_per_seq() const { + return cparams.n_ctx / cparams.n_seq_max; +} + +uint32_t llama_context::n_batch() const { + return cparams.n_batch; +} + +uint32_t llama_context::n_ubatch() const { + return cparams.n_ubatch; +} + +uint32_t llama_context::n_seq_max() const { + return cparams.n_seq_max; +} + +uint32_t llama_context::n_threads() const { + return cparams.n_threads; +} + +uint32_t llama_context::n_threads_batch() const { + return cparams.n_threads_batch; +} + +llama_kv_cache * llama_context::get_kv_self() { + return kv_self.get(); +} + +const llama_kv_cache * llama_context::get_kv_self() const { + return kv_self.get(); +} + +ggml_tensor * llama_context::build_rope_shift( + ggml_context * ctx0, + ggml_tensor * cur, + ggml_tensor * shift, + ggml_tensor * factors, + float freq_base, + float freq_scale, + ggml_backend_buffer * bbuf) const { + const auto & n_ctx_orig = cparams.n_ctx_orig_yarn; + + const auto & yarn_ext_factor = cparams.yarn_ext_factor; + const auto & yarn_attn_factor = cparams.yarn_attn_factor; + const auto & yarn_beta_fast = cparams.yarn_beta_fast; + const auto & yarn_beta_slow = cparams.yarn_beta_slow; + + const auto & hparams = model.hparams; + + const auto & n_rot = hparams.n_rot; + const auto & rope_type = hparams.rope_type; + + ggml_tensor * tmp; + + if (ggml_is_quantized(cur->type)) { + // dequantize to f32 -> RoPE -> quantize back + tmp = ggml_cast(ctx0, cur, GGML_TYPE_F32); + + if (bbuf) { + for (const auto & backend : backends) { + // Figure out which backend KV cache belongs to + if (ggml_backend_supports_buft(backend.get(), ggml_backend_buffer_get_type(bbuf))) { + ggml_backend_sched_set_tensor_backend(sched.get(), tmp, backend.get()); + break; + } + } + } + + tmp = ggml_rope_ext_inplace(ctx0, tmp, + shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow); + + tmp = ggml_cpy(ctx0, tmp, cur); + } else { + // we rotate only the first n_rot dimensions + tmp = ggml_rope_ext_inplace(ctx0, cur, + shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow); + } + + return tmp; +} + +class llm_graph_input_k_shift : public llm_graph_input_i { +public: + llm_graph_input_k_shift(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {} + virtual ~llm_graph_input_k_shift() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * k_shift; // I32 [kv_size] + + const llama_kv_cache_unified * kv_self; +}; + +void llm_graph_input_k_shift::set_input(const llama_ubatch * ubatch) { + GGML_UNUSED(ubatch); + + if (k_shift) { + assert(ggml_backend_buffer_is_host(k_shift->buffer)); + + int32_t * data = (int32_t *) k_shift->data; + + for (uint32_t i = 0; i < kv_self->size; ++i) { + data[i] = kv_self->cells[i].delta; + } + } +} + +llm_graph_result_ptr llama_context::build_kv_self_shift( + ggml_context * ctx0, + ggml_cgraph * gf) const { + auto res = std::make_unique(); + + const auto & hparams = model.hparams; + + const auto & n_layer = hparams.n_layer; + + const auto & n_embd_head_k = hparams.n_embd_head_k; + //const auto & n_embd_head_v = hparams.n_embd_head_v; + + //GGML_ASSERT(kv_self->size == n_ctx); + + auto inp = std::make_unique(kv_self.get()); + + inp->k_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, cparams.n_ctx); + ggml_set_input(inp->k_shift); + + for (uint32_t il = 0; il < n_layer; ++il) { + const int64_t n_head_kv = hparams.n_head_kv(il); + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); + + float freq_base_l = cparams.rope_freq_base; + float freq_scale_l = cparams.rope_freq_scale; + + // TODO: improve + if (model.arch == LLM_ARCH_GEMMA3) { + const bool is_sliding = hparams.is_sliding(il); + + freq_base_l = is_sliding ? 10000.0f : cparams.rope_freq_base; + freq_scale_l = is_sliding ? 1.0f : cparams.rope_freq_scale; + } + + ggml_tensor * rope_factors = kv_self->cbs.get_rope_factors(n_ctx_per_seq(), il); + + ggml_tensor * k = + ggml_view_3d(ctx0, kv_self->k_l[il], + n_embd_head_k, n_head_kv, kv_self->size, + ggml_row_size(kv_self->k_l[il]->type, n_embd_head_k), + ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa), + 0); + + ggml_tensor * cur = build_rope_shift(ctx0, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l, kv_self->k_l[il]->buffer); + + ggml_build_forward_expand(gf, cur); + } + + res->add_input(std::move(inp)); + + return res; +} + +llm_graph_result_ptr llama_context::build_kv_self_defrag( + ggml_context * ctx0, + ggml_cgraph * gf) const { + auto res = std::make_unique(); + + const auto & hparams = model.hparams; + + const auto & ids = kv_self->defrag_info.ids; + +#if 0 + // CPU defrag // - // set input data + // TODO: optimizations are possible: + // - multiple threads + // - avoid copying to the host memory when already there + // + // likely not worth the effort, as we have ggml_graph based defrag // - const auto & hparams = lctx.model.hparams; - const auto & cparams = lctx.cparams; - const auto & kv_self = lctx.kv_self; + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(); - if (ubatch.token) { - const int64_t n_tokens = ubatch.n_tokens; + const uint32_t kv_size = size; - ggml_backend_tensor_set(lctx.inp_tokens, ubatch.token, 0, n_tokens*ggml_element_size(lctx.inp_tokens)); - } + std::vector buf_k; + std::vector buf_v; - if (ubatch.embd) { - const int64_t n_embd = hparams.n_embd; - const int64_t n_tokens = ubatch.n_tokens; + for (uint32_t il = 0; il < n_layer; ++il) { + const size_t k_size_row = ggml_row_size(k_l[il]->type, n_embd_k_gqa); + const size_t k_size = ggml_row_size(k_l[il]->type, n_embd_k_gqa*kv_size); - ggml_backend_tensor_set(lctx.inp_embd, ubatch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd)); - } + const size_t v_size_el = ggml_type_size(v_l[il]->type); + const size_t v_size = ggml_row_size (v_l[il]->type, n_embd_v_gqa*kv_size); - if (ubatch.pos && lctx.inp_pos) { - const int64_t n_tokens = ubatch.n_tokens; - auto n_pos = lctx.n_pos_per_token; - ggml_backend_tensor_set(lctx.inp_pos, ubatch.pos, 0, n_tokens*n_pos*ggml_element_size(lctx.inp_pos)); - } + buf_k.resize(k_size); + buf_v.resize(v_size); - if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) { - //GGML_ASSERT(lctx.inp_out_ids && "every model that can must skip unused outputs"); + ggml_backend_tensor_get(k_l[il], buf_k.data(), 0, buf_k.size()); + ggml_backend_tensor_get(v_l[il], buf_v.data(), 0, buf_v.size()); - if (!lctx.inp_out_ids) { - LLAMA_LOG_WARN("%s: 'lctx.inp_out_ids' is not created\n", __func__); - } else { - const int64_t n_tokens = ubatch.n_tokens; + // batch move [i, i+nm) to [id, id+nm) + // note: cells can move only to a lower index + for (uint32_t i = 0; i < n_kv; ++i) { + const uint32_t id = ids[i]; - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_out_ids->buffer)); - int32_t * data = (int32_t *) lctx.inp_out_ids->data; + if (i == id || id == n_kv) { + continue; + } - if (lctx.n_outputs == n_tokens) { - for (int i = 0; i < n_tokens; ++i) { - data[i] = i; + uint32_t nm = 1; + + while (i + nm < n_kv && ids[i + nm] == id + nm) { + nm++; + } + + // move keys + { + const int64_t os = i*k_size_row; + const int64_t od = id*k_size_row; + + memcpy(buf_k.data() + od, buf_k.data() + os, nm*k_size_row); + } + + // move values (note: they are transposed) + { + const int64_t os = i; + const int64_t od = id; + + for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { + memcpy(buf_v.data() + (od + j*kv_size)*v_size_el, buf_v.data() + (os + j*kv_size)*v_size_el, nm*v_size_el); } - } else if (ubatch.output) { - int32_t n_outputs = 0; - for (int i = 0; i < n_tokens; ++i) { - if (ubatch.output[i]) { - data[n_outputs++] = i; - } - } - // the graph needs to have been passed the correct number of outputs - GGML_ASSERT(lctx.n_outputs == n_outputs); - } else if (lctx.n_outputs == 1) { - // only keep last output - data[0] = n_tokens - 1; + } + + i += nm - 1; + } + + ggml_backend_tensor_set(k_l[il], buf_k.data(), 0, buf_k.size()); + ggml_backend_tensor_set(v_l[il], buf_v.data(), 0, buf_v.size()); + } +#else + for (uint32_t i = 0; i < ids.size(); ++i) { + const uint32_t id = ids[i]; + + if (i == id || id == ids.size()) { + continue; + } + + uint32_t nm = 1; + + while (i + nm < ids.size() && ids[i + nm] == id + nm) { + nm++; + } + + for (uint32_t il = 0; il < hparams.n_layer; ++il) { // NOLINT + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); + const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); + + ggml_tensor * view_k_src = ggml_view_2d(ctx0, kv_self->k_l[il], + n_embd_k_gqa, nm, + ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa), + ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa*i)); + + ggml_tensor * view_k_dst = ggml_view_2d(ctx0, kv_self->k_l[il], + n_embd_k_gqa, nm, + ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa), + ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa*id)); + + ggml_tensor * view_v_src; + ggml_tensor * view_v_dst; + + if (cparams.flash_attn) { + // NOTE: the V cache is not transposed when using flash attention + view_v_src = ggml_view_2d(ctx0, kv_self->v_l[il], + n_embd_v_gqa, nm, + ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa), + ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa*i)); + + view_v_dst = ggml_view_2d(ctx0, kv_self->v_l[il], + n_embd_v_gqa, nm, + ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa), + ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa*id)); } else { - GGML_ASSERT(lctx.n_outputs == 0); + view_v_src = ggml_view_2d(ctx0, kv_self->v_l[il], + nm, n_embd_v_gqa, + ggml_row_size(kv_self->v_l[il]->type, kv_self->size), + ggml_row_size(kv_self->v_l[il]->type, i)); + + view_v_dst = ggml_view_2d(ctx0, kv_self->v_l[il], + nm, n_embd_v_gqa, + ggml_row_size(kv_self->v_l[il]->type, kv_self->size), + ggml_row_size(kv_self->v_l[il]->type, id)); + } + + ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_k_src, view_k_dst)); + ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_v_src, view_v_dst)); + } + + i += nm - 1; + } + + //LLAMA_LOG_INFO("gf->n_nodes = %d\n", gf->n_nodes); +#endif + + return res; +} + +void llama_context::kv_self_update() { + auto & kv = kv_self; + + bool need_reserve = false; + + if (kv->has_shift) { + if (!kv->get_can_shift()) { + GGML_ABORT("The current context does not support K-shift"); + } + + LLAMA_LOG_DEBUG("%s: applying K-shift\n", __func__); + + // apply K-shift if needed + if (model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE) { + ggml_backend_sched_reset(sched.get()); + + auto * gf = graph_init(); + + auto res = build_kv_self_shift(ctx_compute.get(), gf); + + ggml_backend_sched_alloc_graph(sched.get(), gf); + + res->set_inputs(nullptr); + + graph_compute(gf, false); + + need_reserve = true; + } + + { + kv->has_shift = false; + + for (uint32_t i = 0; i < kv->size; ++i) { + kv->cells[i].delta = 0; } } } - GGML_ASSERT( - // (!a || b) is a logical implication (a -> b) - // !hparams.causal_attn -> !cparams.causal_attn - (hparams.causal_attn || !cparams.causal_attn) && - "causal attention is not supported by this model" - ); + // defragment the KV cache if needed + if (kv->do_defrag) { + LLAMA_LOG_DEBUG("%s: defragmenting KV cache\n", __func__); - if (lctx.inp_KQ_mask || lctx.inp_KQ_mask_swa) { - // NOTE: hparams.causal_attn indicates the model is capable of generation and uses the kv cache. - if (cparams.causal_attn && !lctx.is_encoding) { - const int64_t n_kv = kv_self.n; - const int64_t n_tokens = ubatch.n_tokens; - const int64_t n_seq_tokens = ubatch.n_seq_tokens; - const int64_t n_seqs = ubatch.n_seqs; + if (kv->defrag_prepare(graph_max_nodes())) { + ggml_backend_sched_reset(sched.get()); + auto * gf = graph_init(); - float * data = nullptr; - float * data_swa = nullptr; + auto res = build_kv_self_defrag(ctx_compute.get(), gf); - if (lctx.inp_KQ_mask) { - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask->buffer)); - data = (float *) lctx.inp_KQ_mask->data; - } + ggml_backend_sched_alloc_graph(sched.get(), gf); - if (lctx.inp_KQ_mask_swa) { - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask_swa->buffer)); - data_swa = (float *) lctx.inp_KQ_mask_swa->data; - } + res->set_inputs(nullptr); - // For causal attention, use only the previous KV cells - // of the correct sequence for each token of the ubatch. - // It's assumed that if a token in the batch has multiple sequences, they are equivalent. - for (int h = 0; h < 1; ++h) { - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; + graph_compute(gf, false); - for (int j = 0; j < n_seq_tokens; ++j) { - const llama_pos pos = ubatch.pos[s*n_seq_tokens + j]; - - for (int i = 0; i < n_kv; ++i) { - float f; - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - f = -INFINITY; - } else { - if (hparams.use_alibi) { - f = -std::abs(kv_self.cells[i].pos - pos); - } else { - f = 0.0f; - } - } - - if (data) { - data[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f; - } - - // may need to cut off old tokens for sliding window - if (data_swa) { - if (pos - kv_self.cells[i].pos >= (int32_t)hparams.n_swa) { - f = -INFINITY; - } - data_swa[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f; - } - } - } - } - - if (data) { - for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { - for (int j = 0; j < n_kv; ++j) { - data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY; - } - } - } - - if (data_swa) { - for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { - for (int j = 0; j < n_kv; ++j) { - data_swa[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY; - } - } - } - } - } else { - const int64_t n_tokens = ubatch.n_tokens; - const int64_t n_seq_tokens = ubatch.n_seq_tokens; - const int64_t n_seqs = ubatch.n_seqs; - // when using kv cache, the mask needs to match the kv cache size - const int64_t n_stride = hparams.causal_attn && !lctx.is_encoding ? kv_self.n : n_tokens; - - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask->buffer)); - - float * data = (float *) lctx.inp_KQ_mask->data; - - for (int h = 0; h < 1; ++h) { - for (int s1 = 0; s1 < n_seqs; ++s1) { - const llama_seq_id seq_id = ubatch.seq_id[s1][0]; - - for (int j = 0; j < n_seq_tokens; ++j) { - const int32_t tj = s1*n_seq_tokens + j; - - for (int s0 = 0; s0 < n_seqs; ++s0) { - for (int i = 0; i < n_seq_tokens; ++i) { - const int32_t ti = s0*n_seq_tokens + i; - float f = -INFINITY; - - for (int s = 0; s < ubatch.n_seq_id[s0]; ++s) { - if (ubatch.seq_id[s0][s] == seq_id) { - if (hparams.use_alibi) { - f = -std::abs(ubatch.pos[ti] - ubatch.pos[tj]); - } else { - f = 0.0f; - } - break; - } - } - - data[h*(n_tokens*n_tokens) + tj*n_stride + ti] = f; - } - } - - for (int i = n_tokens; i < n_stride; ++i) { - data[h*(n_tokens*n_tokens) + tj*n_stride + i] = -INFINITY; - } - } - } - } + need_reserve = true; } + + kv->do_defrag = false; } - if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) { - const int64_t n_tokens = ubatch.n_tokens; - const int64_t n_seq_tokens = ubatch.n_seq_tokens; - const int64_t n_seqs = ubatch.n_seqs; + // reserve a worst case graph if needed + if (need_reserve) { + LLAMA_LOG_DEBUG("%s: reserving a worst case graph\n", __func__); - GGML_ASSERT(lctx.inp_mean); - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_mean->buffer)); + // build worst-case graph + uint32_t n_seqs = 1; // TODO: worst-case number of sequences + uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch); - float * data = (float *) lctx.inp_mean->data; - memset(lctx.inp_mean->data, 0, n_tokens * n_tokens * ggml_element_size(lctx.inp_mean)); + // simulate full KV cache + kv_self->n = kv_self->size; - std::vector sum(n_tokens, 0); + llama_token token = model.vocab.token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph + llama_ubatch ubatch = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr}; - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; + auto * gf = graph_init(); + graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT); - // TODO: adapt limits to n_seqs when ubatch.equal_seqs is true - GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN"); - - sum[seq_id] += ubatch.n_seq_tokens; - } - - std::vector div(n_tokens, 0.0f); - for (int i = 0; i < n_tokens; ++i) { - const uint64_t s = sum[i]; - if (s > 0) { - div[i] = 1.0f/float(s); - } - } - - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; - - for (int i = 0; i < n_seq_tokens; ++i) { - data[seq_id*n_tokens + s*n_seq_tokens + i] = div[seq_id]; - } - } - } - - if (cparams.embeddings && ( - cparams.pooling_type == LLAMA_POOLING_TYPE_CLS || - cparams.pooling_type == LLAMA_POOLING_TYPE_RANK)) { - const int64_t n_tokens = ubatch.n_tokens; - const int64_t n_seq_tokens = ubatch.n_seq_tokens; - const int64_t n_seqs = ubatch.n_seqs; - - GGML_ASSERT(lctx.inp_cls); - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer)); - - uint32_t * data = (uint32_t *) lctx.inp_cls->data; - memset(lctx.inp_cls->data, 0, n_tokens * ggml_element_size(lctx.inp_cls)); - - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; - - // TODO: adapt limits to n_seqs when ubatch.equal_seqs is true - GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS or RANK"); - - for (int i = 0; i < n_seq_tokens; ++i) { - const llama_pos pos = ubatch.pos[s*n_seq_tokens + i]; - - if (pos == 0) { - data[seq_id] = s*n_seq_tokens + i; - } - } - } - } - - if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_LAST) { - const int64_t n_tokens = ubatch.n_tokens; - const int64_t n_seq_tokens = ubatch.n_seq_tokens; - const int64_t n_seqs = ubatch.n_seqs; - - GGML_ASSERT(lctx.inp_cls); - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer)); - - uint32_t * data = (uint32_t *) lctx.inp_cls->data; - memset(lctx.inp_cls->data, 0, n_tokens * ggml_element_size(lctx.inp_cls)); - - std::vector last_pos(n_tokens, -1); - std::vector last_row(n_tokens, -1); - - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; - - // TODO: adapt limits to n_seqs when ubatch.equal_seqs is true - GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == LAST"); - - for (int i = 0; i < n_seq_tokens; ++i) { - const llama_pos pos = ubatch.pos[s*n_seq_tokens + i]; - - if (pos >= last_pos[seq_id]) { - last_pos[seq_id] = pos; - last_row[seq_id] = s*n_seq_tokens + i; - } - } - } - - for (int i = 0; i < n_tokens; ++i) { - if (last_row[i] >= 0) { - data[i] = last_row[i]; - } - } - } - - if (kv_self.recurrent) { - const int64_t n_kv = kv_self.n; - - if (lctx.inp_s_mask) { - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_s_mask->buffer)); - float * data = (float *) lctx.inp_s_mask->data; - - // clear unused states - for (int i = 0; i < n_kv; ++i) { - const uint32_t cell_id = i + kv_self.head; - llama_kv_cell & kv_cell = lctx.kv_self.cells[cell_id]; - - data[i] = (float) (kv_cell.src >= 0); - - // only clear once - if (kv_cell.src < 0) { - kv_cell.src = cell_id; - } - } - } - - if (lctx.inp_s_copy) { - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_s_copy->buffer)); - int32_t * data = (int32_t *) lctx.inp_s_copy->data; - - // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n - for (uint32_t i = 0; i < n_kv; ++i) { - const uint32_t cell_id = i + kv_self.head; - llama_kv_cell & kv_cell = lctx.kv_self.cells[cell_id]; - - // prevent out-of-bound sources - if (kv_cell.src < 0 || (uint32_t) kv_cell.src >= kv_self.size) { - kv_cell.src = cell_id; - } - - data[i] = kv_cell.src; - - // ensure copy only happens once - if (kv_cell.src != (int32_t) cell_id) { - kv_cell.src = cell_id; - } - } - } - } - - if (lctx.inp_pos_bucket) { - const int64_t n_tokens = ubatch.n_tokens; - - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_pos_bucket->buffer)); - GGML_ASSERT(!ubatch.equal_seqs); // TODO: use ubatch.n_seqs instead of failing - - int32_t * data = (int32_t *) lctx.inp_pos_bucket->data; - - if (!lctx.is_encoding) { - const int64_t n_kv = kv_self.n; - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - for (int i = 0; i < n_kv; ++i) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(lctx.kv_self.cells[i].pos, ubatch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding); - } - } - } - } else { - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - for (int i = 0; i < n_tokens; ++i) { - data[h*(n_tokens*n_tokens) + j*n_tokens + i] = llama_relative_position_bucket(ubatch.pos[i], ubatch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding); - } - } - } - } - } - - if (!lctx.is_encoding && lctx.inp_embd_enc) { - assert(lctx.inp_embd_enc->type == GGML_TYPE_F32); - assert((size_t) ggml_nelements(lctx.inp_embd_enc) == lctx.embd_enc.size()); - - ggml_backend_tensor_set(lctx.inp_embd_enc, lctx.embd_enc.data(), 0, ggml_nbytes(lctx.inp_embd_enc)); - } - - if (!lctx.is_encoding && lctx.inp_KQ_mask_cross) { - const int64_t n_output_enc = lctx.embd_enc.size() / hparams.n_embd; - const int64_t n_tokens = ubatch.n_tokens; - - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask_cross->buffer)); - GGML_ASSERT(!ubatch.equal_seqs); // TODO: use ubatch.n_seqs instead of failing - - float * data = (float *) lctx.inp_KQ_mask_cross->data; - - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - for (int i = 0; i < n_output_enc; ++i) { - float f = -INFINITY; - for (int s = 0; s < ubatch.n_seq_id[j]; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[j][s]; - if (lctx.seq_ids_enc[i].find(seq_id) != lctx.seq_ids_enc[i].end()) { - f = 0.0f; - } - } - data[h*(n_output_enc*n_tokens) + j*n_output_enc + i] = f; - } - } - - for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { - for (int j = 0; j < n_output_enc; ++j) { - data[h*(n_output_enc*n_tokens) + i*n_output_enc + j] = -INFINITY; - } - } + // initialize scheduler with the worst-case graph + ggml_backend_sched_reset(sched.get()); + if (!ggml_backend_sched_reserve(sched.get(), gf)) { + LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__); } } } -// llama output +enum llama_pooling_type llama_context::pooling_type() const { + return cparams.pooling_type; +} -size_t llama_output_reserve(struct llama_context & lctx, size_t n_outputs) { - const auto & cparams = lctx.cparams; - const auto & hparams = lctx.model.hparams; - const auto & vocab = lctx.model.vocab; +float * llama_context::get_logits() { + // reorder logits for backward compatibility + output_reorder(); - const size_t n_outputs_max = std::max(n_outputs, (size_t) cparams.n_seq_max); + return logits; +} + +float * llama_context::get_logits_ith(int32_t i) { + int32_t j = -1; + + try { + if (logits == nullptr) { + throw std::runtime_error("no logits"); + } + + if (i < 0) { + j = n_outputs + i; + if (j < 0) { + throw std::runtime_error(format("negative index out of range [0, %d)", n_outputs)); + } + } else if ((size_t) i >= output_ids.size()) { + throw std::runtime_error(format("out of range [0, %zu)", output_ids.size())); + } else { + j = output_ids[i]; + } + + if (j < 0) { + throw std::runtime_error(format("batch.logits[%d] != true", i)); + } + if (j >= n_outputs) { + // This should not happen + throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, n_outputs)); + } + + return logits + j*model.vocab.n_tokens(); + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: invalid logits id %d, reason: %s\n", __func__, i, err.what()); +#ifndef NDEBUG + GGML_ABORT("fatal error"); +#else + return nullptr; +#endif + } +} + +float * llama_context::get_embeddings() { + // reorder embeddings for backward compatibility + output_reorder(); + + return embd; +} + +float * llama_context::get_embeddings_ith(int32_t i) { + int32_t j = -1; + + try { + if (embd == nullptr) { + throw std::runtime_error("no embeddings"); + } + + if (i < 0) { + j = n_outputs + i; + if (j < 0) { + throw std::runtime_error(format("negative index out of range [0, %d)", n_outputs)); + } + } else if ((size_t) i >= output_ids.size()) { + throw std::runtime_error(format("out of range [0, %zu)", output_ids.size())); + } else { + j = output_ids[i]; + } + + if (j < 0) { + throw std::runtime_error(format("batch.logits[%d] != true", i)); + } + if (j >= n_outputs) { + // This should not happen + throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, n_outputs)); + } + + return embd + j*model.hparams.n_embd; + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: invalid embeddings id %d, reason: %s\n", __func__, i, err.what()); +#ifndef NDEBUG + GGML_ABORT("fatal error"); +#else + return nullptr; +#endif + } +} + +float * llama_context::get_embeddings_seq(llama_seq_id seq_id) { + auto it = embd_seq.find(seq_id); + if (it == embd_seq.end()) { + return nullptr; + } + + return it->second.data(); +} + +void llama_context::attach_threadpool( + ggml_threadpool_t threadpool, + ggml_threadpool_t threadpool_batch) { + LLAMA_LOG_DEBUG("%s: call\n", __func__); + + this->threadpool = threadpool; + this->threadpool_batch = threadpool_batch ? threadpool_batch : threadpool; +} + +void llama_context::detach_threadpool() { + LLAMA_LOG_DEBUG("%s: call\n", __func__); + + this->threadpool = nullptr; + this->threadpool_batch = nullptr; +} + +void llama_context::set_n_threads(int32_t n_threads, int32_t n_threads_batch) { + LLAMA_LOG_DEBUG("%s: n_threads = %d, n_threads_batch = %d\n", __func__, n_threads, n_threads_batch); + + cparams.n_threads = n_threads; + cparams.n_threads_batch = n_threads_batch; +} + +void llama_context::set_abort_callback(bool (*abort_callback)(void * data), void * abort_callback_data) { + LLAMA_LOG_DEBUG("%s: call\n", __func__); + + this->abort_callback = abort_callback; + this->abort_callback_data = abort_callback_data; + + for (auto & backend : backends) { + auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend.get())); + auto * set_abort_callback_fn = (ggml_backend_set_abort_callback_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback"); + if (set_abort_callback_fn) { + set_abort_callback_fn(backend.get(), this->abort_callback, this->abort_callback_data); + } + } +} + +void llama_context::set_embeddings(bool value) { + LLAMA_LOG_DEBUG("%s: value = %d\n", __func__, value); + + cparams.embeddings = value; +} + +void llama_context::set_causal_attn(bool value) { + LLAMA_LOG_DEBUG("%s: value = %d\n", __func__, value); + + cparams.causal_attn = value; +} + +void llama_context::set_adapter_lora( + llama_adapter_lora * adapter, + float scale) { + LLAMA_LOG_DEBUG("%s: adapter = %p, scale = %f\n", __func__, (void *) adapter, scale); + + loras[adapter] = scale; +} + +bool llama_context::rm_adapter_lora( + llama_adapter_lora * adapter) { + LLAMA_LOG_DEBUG("%s: adapter = %p\n", __func__, (void *) adapter); + + auto pos = loras.find(adapter); + if (pos != loras.end()) { + loras.erase(pos); + return true; + } + + return false; +} + +void llama_context::clear_adapter_lora() { + LLAMA_LOG_DEBUG("%s: call\n", __func__); + + loras.clear(); +} + +bool llama_context::apply_adapter_cvec( + const float * data, + size_t len, + int32_t n_embd, + int32_t il_start, + int32_t il_end) { + LLAMA_LOG_DEBUG("%s: il_start = %d, il_end = %d\n", __func__, il_start, il_end); + + return cvec.apply(model, data, len, n_embd, il_start, il_end); +} + +int llama_context::encode(llama_batch & inp_batch) { + if (inp_batch.n_tokens == 0) { + LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__); + return -1; + } + + // temporary allocate memory for the input batch if needed + // TODO: this is incorrect for multiple sequences because pos_max() is the maximum across all sequences + llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : kv_self->pos_max() + 1); + + const llama_batch & batch = batch_allocr.batch; + const int32_t n_tokens = batch.n_tokens; + + const auto & hparams = model.hparams; + + GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT + + if (batch.token) { + for (int32_t i = 0; i < n_tokens; ++i) { + if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= model.vocab.n_tokens()) { + LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]); + return -1; + } + } + } + + // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot + GGML_ASSERT(cparams.n_ubatch >= (uint32_t) n_tokens && "encoder requires n_ubatch >= n_tokens"); + + if (t_compute_start_us == 0) { + t_compute_start_us = ggml_time_us(); + } + + n_queued_tokens += n_tokens; + + const int64_t n_embd = hparams.n_embd; + + sbatch.from_batch(batch, n_embd, /* simple_split */ true, /* logits_all */ true); + + const llama_ubatch ubatch = sbatch.split_simple(n_tokens); + + // reserve output buffer + if (output_reserve(n_tokens) < n_tokens) { + LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_tokens); + return -2; + }; + + for (int32_t i = 0; i < n_tokens; ++i) { + output_ids[i] = i; + } + + n_outputs = n_tokens; + + //batch_manager->prepare(ubatch); + + ggml_backend_sched_reset(sched.get()); + ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data); + + auto * gf = graph_init(); + auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_ENCODER); + + ggml_backend_sched_alloc_graph(sched.get(), gf); + + res->set_inputs(&ubatch); + + const auto compute_status = graph_compute(gf, n_tokens > 1); + switch (compute_status) { + case GGML_STATUS_SUCCESS: + break; + case GGML_STATUS_ABORTED: + return 2; + case GGML_STATUS_ALLOC_FAILED: + return -2; + case GGML_STATUS_FAILED: + default: + return -3; + } + + auto * t_embd = res->get_embd_pooled() ? res->get_embd_pooled() : res->get_embd(); + + // extract embeddings + if (t_embd) { + ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(sched.get(), t_embd); + GGML_ASSERT(backend_embd != nullptr); + + GGML_ASSERT(embd != nullptr); + + switch (cparams.pooling_type) { + case LLAMA_POOLING_TYPE_NONE: + { + // extract token embeddings + GGML_ASSERT(n_tokens*n_embd <= (int64_t) embd_size); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd, 0, n_tokens*n_embd*sizeof(float)); + } break; + case LLAMA_POOLING_TYPE_MEAN: + case LLAMA_POOLING_TYPE_CLS: + case LLAMA_POOLING_TYPE_LAST: + { + // extract sequence embeddings + auto & embd_seq_out = embd_seq; + embd_seq_out.clear(); + + GGML_ASSERT(!ubatch.equal_seqs); // TODO: handle equal splits + + for (int32_t i = 0; i < n_tokens; i++) { + const llama_seq_id seq_id = ubatch.seq_id[i][0]; + if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { + continue; + } + embd_seq_out[seq_id].resize(n_embd); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float)); + } + } break; + case LLAMA_POOLING_TYPE_RANK: + { + // TODO: this likely should be the same logic as in llama_decoder_internal, but better to + // wait for an encoder model that requires this pooling type in order to test it + // https://github.com/ggerganov/llama.cpp/pull/9510 + GGML_ABORT("RANK pooling not implemented yet"); + } + case LLAMA_POOLING_TYPE_UNSPECIFIED: + { + GGML_ABORT("unknown pooling type"); + } + } + } + + // Reset state for the next token before backend sync, to allow the CPU activities in the reset to + // overlap with device computation. + ggml_backend_sched_reset(sched.get()); + + // TODO: hacky solution + if (model.arch == LLM_ARCH_T5 && t_embd) { + //cross.t_embd = t_embd; + + cross.n_embd = t_embd->ne[0]; + cross.n_enc = t_embd->ne[1]; + cross.v_embd.resize(cross.n_embd*cross.n_enc); + memcpy(cross.v_embd.data(), embd, ggml_nbytes(t_embd)); + + // remember the sequence ids used during the encoding - needed for cross attention later + cross.seq_ids_enc.resize(n_tokens); + for (int32_t i = 0; i < n_tokens; i++) { + for (int s = 0; s < ubatch.n_seq_id[i]; s++) { + llama_seq_id seq_id = ubatch.seq_id[i][s]; + cross.seq_ids_enc[i].insert(seq_id); + } + } + } + + return 0; +} + +int llama_context::decode(llama_batch & inp_batch) { + if (inp_batch.n_tokens == 0) { + LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__); + return -1; + } + + // temporary allocate memory for the input batch if needed + // TODO: this is incorrect for multiple sequences because pos_max() is the maximum across all sequences + llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : kv_self->pos_max() + 1); + + const llama_batch & batch = batch_allocr.batch; + + const auto & vocab = model.vocab; + const auto & hparams = model.hparams; + + const int32_t n_vocab = vocab.n_tokens(); + + const int64_t n_tokens_all = batch.n_tokens; + const int64_t n_embd = hparams.n_embd; + + // TODO: remove this stuff + class batch_guard { + public: + batch_guard(llama_kv_cache_unified & kv_self) : kv_slot_restorer(kv_self) { + } + + ~batch_guard() { + if (!is_done) { + kv_slot_restorer.restore(); + } + } + + void done() { + is_done = true; + } + + void save(const llama_kv_cache_slot_info & slot_info) { + kv_slot_restorer.save(slot_info); + } + + private: + bool is_done = false; + + llama_kv_slot_restorer kv_slot_restorer; + }; + + batch_guard bg(*kv_self); + + GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT + + if (batch.token) { + for (int64_t i = 0; i < n_tokens_all; ++i) { + if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= model.vocab.n_tokens()) { + LLAMA_LOG_ERROR("%s: invalid token[%" PRId64 "] = %d\n", __func__, i, batch.token[i]); + throw std::runtime_error("invalid token"); + } + } + } + + GGML_ASSERT(n_tokens_all <= cparams.n_batch); + + GGML_ASSERT((cparams.causal_attn || cparams.n_ubatch >= n_tokens_all) && "non-causal attention requires n_ubatch >= n_tokens"); + + if (t_compute_start_us == 0) { + t_compute_start_us = ggml_time_us(); + } + n_queued_tokens += n_tokens_all; + + // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens + const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE; + + embd_seq.clear(); + + int64_t n_outputs_all = 0; + + // count outputs + if (batch.logits && !embd_pooled) { + for (uint32_t i = 0; i < n_tokens_all; ++i) { + n_outputs_all += batch.logits[i] != 0; + } + } else if (logits_all || embd_pooled) { + n_outputs_all = n_tokens_all; + } else { + // keep last output only + n_outputs_all = 1; + } + + const bool logits_all = n_outputs_all == n_tokens_all; + + sbatch.from_batch(batch, n_embd, + /* simple_split */ !kv_self->recurrent, + /* logits_all */ logits_all); + + // reserve output buffer + if (output_reserve(n_outputs_all) < n_outputs_all) { + LLAMA_LOG_ERROR("%s: could not reserve space for batch with %" PRId64 " outputs\n", __func__, n_outputs_all); + return -2; + }; + + int64_t n_outputs_prev = 0; + + while (sbatch.n_tokens > 0) { + llama_ubatch ubatch = llama_ubatch(); + + const auto & n_ubatch = cparams.n_ubatch; + + if (kv_self->recurrent) { + if (embd_pooled) { + // Pooled embeddings cannot be split across ubatches (yet) + ubatch = sbatch.split_seq(cparams.n_ubatch); + } else { + // recurrent model architectures are easier to implement + // with equal-length sequences + ubatch = sbatch.split_equal(cparams.n_ubatch); + } + } else { + ubatch = sbatch.split_simple(n_ubatch); + } + + // count the outputs in this u_batch + { + int32_t n_outputs_new = 0; + + if (n_outputs_all == n_tokens_all) { + n_outputs_new = ubatch.n_tokens; + } else { + GGML_ASSERT(ubatch.output); + for (uint32_t i = 0; i < ubatch.n_tokens; i++) { + n_outputs_new += (int32_t) (ubatch.output[i] != 0); + } + } + + // needs to happen before the graph is built + n_outputs = n_outputs_new; + } + + // non-causal masks do not use the KV cache + if (hparams.causal_attn) { + kv_self_update(); + + // if we have enough unused cells before the current head -> + // better to start searching from the beginning of the cache, hoping to fill it + if (kv_self->head > kv_self->used + 2*ubatch.n_tokens) { + kv_self->head = 0; + } + + const auto slot_info = kv_self->find_slot(ubatch); + if (!slot_info) { + LLAMA_LOG_ERROR("%s: failed to prepare ubatch\n", __func__); + return -3; + } + + bg.save(slot_info); + + if (!kv_self->recurrent) { + // a heuristic, to avoid attending the full cache if it is not yet utilized + // after enough generations, the benefit from this heuristic disappears + // if we start defragmenting the cache, the benefit from this will be more important + const uint32_t pad = kv_self->get_padding(cparams); + kv_self->n = std::min(kv_self->size, std::max(pad, GGML_PAD(kv_self->cell_max(), pad))); + } + } + + //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self->n, kv_self->used, kv_self->head); + + ggml_backend_sched_reset(sched.get()); + ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data); + + auto * gf = graph_init(); + auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DECODER); + + // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs); + + ggml_backend_sched_alloc_graph(sched.get(), gf); + + res->set_inputs(&ubatch); + + const auto compute_status = graph_compute(gf, ubatch.n_tokens > 1); + if (compute_status != GGML_STATUS_SUCCESS) { + switch (compute_status) { + case GGML_STATUS_ABORTED: + return 2; + case GGML_STATUS_ALLOC_FAILED: + return -2; + case GGML_STATUS_FAILED: + default: + return -3; + } + } + + // update the kv ring buffer + { + kv_self->head += ubatch.n_tokens; + + // Ensure kv cache head points to a valid index. + if (kv_self->head >= kv_self->size) { + kv_self->head = 0; + } + } + + // plot the computation graph in dot format (for debugging purposes) + //if (n_past%100 == 0) { + // ggml_graph_dump_dot(gf, NULL, "llama.dot"); + //} + + auto * t_logits = cparams.embeddings ? nullptr : res->get_logits(); + auto * t_embd = cparams.embeddings ? res->get_embd() : nullptr; + + if (t_embd && res->get_embd_pooled()) { + t_embd = res->get_embd_pooled(); + } + + // extract logits + if (t_logits && n_outputs > 0) { + ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(sched.get(), t_logits); + GGML_ASSERT(backend_res != nullptr); + GGML_ASSERT(logits != nullptr); + + float * logits_out = logits + n_outputs_prev*n_vocab; + + if (n_outputs) { + GGML_ASSERT( n_outputs_prev + n_outputs <= n_outputs_all); + GGML_ASSERT((n_outputs_prev + n_outputs)*n_vocab <= (int64_t) logits_size); + ggml_backend_tensor_get_async(backend_res, t_logits, logits_out, 0, n_outputs*n_vocab*sizeof(float)); + } + } + + // extract embeddings + if (t_embd && n_outputs > 0) { + ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(sched.get(), t_embd); + GGML_ASSERT(backend_embd != nullptr); + + switch (cparams.pooling_type) { + case LLAMA_POOLING_TYPE_NONE: + { + // extract token embeddings + GGML_ASSERT(embd != nullptr); + float * embd_out = embd + n_outputs_prev*n_embd; + + if (n_outputs) { + GGML_ASSERT( n_outputs_prev + n_outputs <= n_outputs_all); + GGML_ASSERT((n_outputs_prev + n_outputs)*n_embd <= (int64_t) embd_size); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd_out, 0, n_outputs*n_embd*sizeof(float)); + } + } break; + case LLAMA_POOLING_TYPE_MEAN: + case LLAMA_POOLING_TYPE_CLS: + case LLAMA_POOLING_TYPE_LAST: + { + // extract sequence embeddings (cleared before processing each batch) + auto & embd_seq_out = embd_seq; + + for (uint32_t s = 0; s < ubatch.n_seqs; ++s) { + const llama_seq_id seq_id = ubatch.seq_id[s][0]; + if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { + continue; + } + embd_seq_out[seq_id].resize(n_embd); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float)); + } + } break; + case LLAMA_POOLING_TYPE_RANK: + { + // extract the rerank score - a single float per sequence + auto & embd_seq_out = embd_seq; + + for (uint32_t s = 0; s < ubatch.n_seqs; ++s) { + const llama_seq_id seq_id = ubatch.seq_id[s][0]; + if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { + continue; + } + embd_seq_out[seq_id].resize(1); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (seq_id)*sizeof(float), sizeof(float)); + } + } break; + case LLAMA_POOLING_TYPE_UNSPECIFIED: + { + GGML_ABORT("unknown pooling type"); + } + } + } + + n_outputs_prev += n_outputs; + } + + // finalize the batch processing + bg.done(); + + // set output mappings + { + bool sorted_output = true; + + GGML_ASSERT(sbatch.out_ids.size() == (size_t) n_outputs_all); + + for (int64_t i = 0; i < n_outputs_all; ++i) { + int64_t out_id = sbatch.out_ids[i]; + output_ids[out_id] = i; + if (out_id != i) { + sorted_output = false; + } + } + + if (sorted_output) { + sbatch.out_ids.clear(); + } + } + + // set to total number of outputs in the batch, for use in llama_get_logits_ith + n_outputs = n_outputs_all; + + // wait for the computation to finish (automatically done when obtaining the model output) + //synchronize(); + + // decide if we need to defrag the kv cache + if (cparams.causal_attn && cparams.defrag_thold > 0.0f) { + // - do not defrag small contexts (i.e. < 2048 tokens) + // - count the padding towards the number of used tokens + const float fragmentation = kv_self->n >= 2048 ? std::max(0.0f, 1.0f - float(kv_self->used + kv_self->get_padding(cparams))/float(kv_self->n)) : 0.0f; + + // queue defragmentation for next llama_kv_cache_update + if (fragmentation > cparams.defrag_thold) { + LLAMA_LOG_DEBUG("%s: fragmentation: %.2f - requesting defrag\n", __func__, fragmentation); + + kv_self->defrag(); + } + } + + // Reset state for the next token before backend sync, to allow the CPU activities in the reset to + // overlap with device computation. + ggml_backend_sched_reset(sched.get()); + + return 0; +} + +// +// output +// + +int32_t llama_context::output_reserve(int32_t n_outputs) { + const auto & hparams = model.hparams; + const auto & vocab = model.vocab; + + const int64_t n_outputs_max = std::max(n_outputs, n_seq_max()); const auto n_batch = cparams.n_batch; const auto n_vocab = vocab.n_tokens(); const auto n_embd = hparams.n_embd; // TODO: use a per-batch flag for logits presence instead - const bool has_logits = !cparams.embeddings; - const bool has_embd = cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE); + bool has_logits = !cparams.embeddings; + bool has_embd = cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE); - const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0; - const size_t embd_size = has_embd ? n_embd*n_outputs_max : 0; - - if (lctx.output_ids.empty()) { - // init, never resized afterwards - lctx.output_ids.resize(n_batch); + // TODO: hacky enc-dec support + if (model.arch == LLM_ARCH_T5) { + has_logits = true; + has_embd = true; } - const size_t prev_size = lctx.buf_output ? ggml_backend_buffer_get_size(lctx.buf_output.get()) : 0; + logits_size = has_logits ? n_vocab*n_outputs_max : 0; + embd_size = has_embd ? n_embd*n_outputs_max : 0; + + if (output_ids.empty()) { + // init, never resized afterwards + output_ids.resize(n_batch); + } + + const size_t prev_size = buf_output ? ggml_backend_buffer_get_size(buf_output.get()) : 0; const size_t new_size = (logits_size + embd_size) * sizeof(float); // alloc only when more than the current capacity is required // TODO: also consider shrinking the buffer - if (!lctx.buf_output || prev_size < new_size) { - if (lctx.buf_output) { + if (!buf_output || prev_size < new_size) { + if (buf_output) { #ifndef NDEBUG // This doesn't happen often, but may be annoying in some cases (like the HellaSwag benchmark) LLAMA_LOG_INFO("%s: reallocating output buffer from size %.02f MiB to %.02f MiB\n", __func__, prev_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0); #endif - lctx.buf_output = nullptr; - lctx.logits = nullptr; - lctx.embd = nullptr; + buf_output = nullptr; + logits = nullptr; + embd = nullptr; } auto * buft = ggml_backend_cpu_buffer_type(); // try to use the host buffer of the device where the output tensor is allocated for faster transfer to system memory - auto * output_dev = lctx.model.dev_output(); + auto * output_dev = model.dev_output(); auto * output_dev_host_buft = output_dev ? ggml_backend_dev_host_buffer_type(output_dev) : nullptr; if (output_dev_host_buft) { buft = output_dev_host_buft; } - lctx.buf_output.reset(ggml_backend_buft_alloc_buffer(buft, new_size)); - if (lctx.buf_output == nullptr) { + buf_output.reset(ggml_backend_buft_alloc_buffer(buft, new_size)); + if (buf_output == nullptr) { LLAMA_LOG_ERROR("%s: failed to allocate output buffer of size %.2f MiB\n", __func__, new_size / (1024.0 * 1024.0)); return 0; } } - float * output_base = (float *) ggml_backend_buffer_get_base(lctx.buf_output.get()); + float * output_base = (float *) ggml_backend_buffer_get_base(buf_output.get()); - lctx.logits = has_logits ? output_base : nullptr; - lctx.embd = has_embd ? output_base + logits_size : nullptr; - - lctx.output_size = n_outputs_max; - lctx.logits_size = logits_size; - lctx.embd_size = embd_size; + logits = has_logits ? output_base : nullptr; + embd = has_embd ? output_base + logits_size : nullptr; // set all ids as invalid (negative) - std::fill(lctx.output_ids.begin(), lctx.output_ids.end(), -1); + std::fill(output_ids.begin(), output_ids.end(), -1); - ggml_backend_buffer_clear(lctx.buf_output.get(), 0); + ggml_backend_buffer_clear(buf_output.get(), 0); - lctx.n_outputs = 0; + this->n_outputs = 0; + this->n_outputs_max = n_outputs_max; return n_outputs_max; } -void llama_output_reorder(struct llama_context & ctx) { - std::vector & out_ids = ctx.sbatch.out_ids; +void llama_context::output_reorder() { + auto & out_ids = sbatch.out_ids; if (!out_ids.empty()) { - const uint32_t n_vocab = ctx.model.vocab.n_tokens(); - const uint32_t n_embd = ctx.model.hparams.n_embd; + const uint32_t n_vocab = model.vocab.n_tokens(); + const uint32_t n_embd = model.hparams.n_embd; - const int32_t n_outputs = ctx.n_outputs; GGML_ASSERT((size_t) n_outputs == out_ids.size()); // TODO: is there something more efficient which also minimizes swaps? @@ -559,842 +1574,156 @@ void llama_output_reorder(struct llama_context & ctx) { } if (j_min == i) { continue; } std::swap(out_ids[i], out_ids[j_min]); - if (ctx.logits_size > 0) { + if (logits_size > 0) { for (uint32_t k = 0; k < n_vocab; k++) { - std::swap(ctx.logits[i*n_vocab + k], ctx.logits[j_min*n_vocab + k]); + std::swap(logits[i*n_vocab + k], logits[j_min*n_vocab + k]); } } - if (ctx.embd_size > 0) { + if (embd_size > 0) { for (uint32_t k = 0; k < n_embd; k++) { - std::swap(ctx.embd[i*n_embd + k], ctx.embd[j_min*n_embd + k]); + std::swap(embd[i*n_embd + k], embd[j_min*n_embd + k]); } } } - std::fill(ctx.output_ids.begin(), ctx.output_ids.end(), -1); + std::fill(output_ids.begin(), output_ids.end(), -1); for (int32_t i = 0; i < n_outputs; ++i) { - ctx.output_ids[out_ids[i]] = i; + output_ids[out_ids[i]] = i; } out_ids.clear(); } } // -// interface implementation +// graph // -void llama_free(struct llama_context * ctx) { - delete ctx; +int32_t llama_context::graph_max_nodes() const { + return std::max(8192, 5*model.n_tensors()); } -uint32_t llama_n_ctx(const struct llama_context * ctx) { - return ctx->cparams.n_ctx; +ggml_cgraph * llama_context::graph_init() { + ggml_init_params params = { + /*.mem_size =*/ buf_compute_meta.size(), + /*.mem_buffer =*/ buf_compute_meta.data(), + /*.no_alloc =*/ true, + }; + + ctx_compute.reset(ggml_init(params)); + + return ggml_new_graph_custom(ctx_compute.get(), graph_max_nodes(), false); } -uint32_t llama_n_batch(const struct llama_context * ctx) { - return ctx->cparams.n_batch; +llm_graph_result_ptr llama_context::graph_build( + ggml_context * ctx, + ggml_cgraph * gf, + const llama_ubatch & ubatch, + llm_graph_type gtype) { + return model.build_graph( + { + /*.ctx =*/ ctx, + /*.arch =*/ model.arch, + /*.hparams =*/ model.hparams, + /*.cparams =*/ cparams, + /*.ubatch =*/ ubatch, + /*.sched =*/ sched.get(), + /*.backend_cpu =*/ backend_cpu, + /*.cvec =*/ &cvec, + /*.loras =*/ &loras, + /*.memory =*/ kv_self.get(), + /*.cross =*/ &cross, + /*.n_outputs =*/ n_outputs, + /*.cb =*/ graph_get_cb(), + }, gf, gtype); } -uint32_t llama_n_ubatch(const struct llama_context * ctx) { - return ctx->cparams.n_ubatch; -} +ggml_status llama_context::graph_compute( + ggml_cgraph * gf, + bool batched) { + int n_threads = batched ? cparams.n_threads_batch : cparams.n_threads; + ggml_threadpool_t tp = batched ? threadpool_batch : threadpool; -uint32_t llama_n_seq_max(const struct llama_context * ctx) { - return ctx->kv_self.size; -} - -const struct llama_model * llama_get_model(const struct llama_context * ctx) { - return &ctx->model; -} - -enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx) { - return ctx->cparams.pooling_type; -} - -void llama_attach_threadpool( - struct llama_context * ctx, - ggml_threadpool_t threadpool, - ggml_threadpool_t threadpool_batch) { - ctx->threadpool = threadpool; - ctx->threadpool_batch = threadpool_batch ? threadpool_batch : threadpool; -} - -void llama_detach_threadpool(struct llama_context * ctx) { - ctx->threadpool = nullptr; - ctx->threadpool_batch = nullptr; -} - -void llama_set_n_threads(struct llama_context * ctx, int32_t n_threads, int32_t n_threads_batch) { - ctx->cparams.n_threads = n_threads; - ctx->cparams.n_threads_batch = n_threads_batch; -} - -int32_t llama_n_threads(struct llama_context * ctx) { - return ctx->cparams.n_threads; -} - -int32_t llama_n_threads_batch(struct llama_context * ctx) { - return ctx->cparams.n_threads_batch; -} - -void llama_set_abort_callback(struct llama_context * ctx, bool (*abort_callback)(void * data), void * abort_callback_data) { - ctx->abort_callback = abort_callback; - ctx->abort_callback_data = abort_callback_data; - - for (auto & backend : ctx->backends) { - auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend.get())); - auto * set_abort_callback_fn = (ggml_backend_set_abort_callback_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback"); - if (set_abort_callback_fn) { - set_abort_callback_fn(backend.get(), ctx->abort_callback, ctx->abort_callback_data); - } - } -} - -void llama_set_embeddings(struct llama_context * ctx, bool embeddings) { - ctx->cparams.embeddings = embeddings; -} - -void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn) { - ctx->cparams.causal_attn = causal_attn; -} - -void llama_synchronize(struct llama_context * ctx) { - ggml_backend_sched_synchronize(ctx->sched.get()); - - // FIXME: if multiple single tokens are evaluated without a synchronization, - // the stats will be added to the prompt evaluation stats - // this should only happen when using batch size 1 to evaluate a batch - - // add the evaluation to the stats - if (ctx->n_queued_tokens == 1) { - if (!ctx->cparams.no_perf) { - ctx->t_eval_us += ggml_time_us() - ctx->t_compute_start_us; - } - ctx->n_eval++; - } else if (ctx->n_queued_tokens > 1) { - if (!ctx->cparams.no_perf) { - ctx->t_p_eval_us += ggml_time_us() - ctx->t_compute_start_us; - } - ctx->n_p_eval += ctx->n_queued_tokens; + if (backend_cpu != nullptr) { + auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend_cpu)); + auto * set_threadpool_fn = (decltype(ggml_backend_cpu_set_threadpool) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool"); + set_threadpool_fn(backend_cpu, tp); } - // get a more accurate load time, upon first eval - if (ctx->n_queued_tokens > 0 && !ctx->has_evaluated_once) { - ctx->t_load_us = ggml_time_us() - ctx->t_start_us; - ctx->has_evaluated_once = true; + // set the number of threads for all the backends + for (const auto & set_n_threads_fn : set_n_threads_fns) { + set_n_threads_fn.second(set_n_threads_fn.first, n_threads); } - ctx->n_queued_tokens = 0; - ctx->t_compute_start_us = 0; + auto status = ggml_backend_sched_graph_compute_async(sched.get(), gf); + if (status != GGML_STATUS_SUCCESS) { + LLAMA_LOG_ERROR("%s: ggml_backend_sched_graph_compute_async failed with error %d\n", __func__, status); + } + + // fprintf(stderr, "splits: %d\n", ggml_backend_sched_get_n_splits(sched)); + + return status; } -float * llama_get_logits(struct llama_context * ctx) { - llama_synchronize(ctx); - - // reorder logits for backward compatibility - // TODO: maybe deprecate this - llama_output_reorder(*ctx); - - return ctx->logits; -} - -float * llama_get_logits_ith(struct llama_context * ctx, int32_t i) { - int32_t j = -1; - - llama_synchronize(ctx); - - try { - if (ctx->logits == nullptr) { - throw std::runtime_error("no logits"); - } - - if (i < 0) { - j = ctx->n_outputs + i; - if (j < 0) { - throw std::runtime_error(format("negative index out of range [0, %d)", ctx->n_outputs)); - } - } else if ((size_t) i >= ctx->output_ids.size()) { - throw std::runtime_error(format("out of range [0, %zu)", ctx->output_ids.size())); +llm_graph_cb llama_context::graph_get_cb() const { + return [&](const llama_ubatch & ubatch, ggml_tensor * cur, const char * name, int il) { + if (il >= 0) { + ggml_format_name(cur, "%s-%d", name, il); } else { - j = ctx->output_ids[i]; + ggml_set_name(cur, name); } - if (j < 0) { - throw std::runtime_error(format("batch.logits[%d] != true", i)); - } - if (j >= ctx->n_outputs) { - // This should not happen - throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, ctx->n_outputs)); - } - - return ctx->logits + j*ctx->model.vocab.n_tokens(); - } catch (const std::exception & err) { - LLAMA_LOG_ERROR("%s: invalid logits id %d, reason: %s\n", __func__, i, err.what()); -#ifndef NDEBUG - GGML_ABORT("fatal error"); -#else - return nullptr; -#endif - } -} - -float * llama_get_embeddings(struct llama_context * ctx) { - llama_synchronize(ctx); - - // reorder embeddings for backward compatibility - // TODO: maybe deprecate this - llama_output_reorder(*ctx); - - return ctx->embd; -} - -float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i) { - int32_t j = -1; - - llama_synchronize(ctx); - - try { - if (ctx->embd == nullptr) { - throw std::runtime_error("no embeddings"); - } - - if (i < 0) { - j = ctx->n_outputs + i; - if (j < 0) { - throw std::runtime_error(format("negative index out of range [0, %d)", ctx->n_outputs)); - } - } else if ((size_t) i >= ctx->output_ids.size()) { - throw std::runtime_error(format("out of range [0, %zu)", ctx->output_ids.size())); - } else { - j = ctx->output_ids[i]; - } - - if (j < 0) { - throw std::runtime_error(format("batch.logits[%d] != true", i)); - } - if (j >= ctx->n_outputs) { - // This should not happen - throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, ctx->n_outputs)); - } - - return ctx->embd + j*ctx->model.hparams.n_embd; - } catch (const std::exception & err) { - LLAMA_LOG_ERROR("%s: invalid embeddings id %d, reason: %s\n", __func__, i, err.what()); -#ifndef NDEBUG - GGML_ABORT("fatal error"); -#else - return nullptr; -#endif - } -} - -float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id) { - llama_synchronize(ctx); - - auto it = ctx->embd_seq.find(seq_id); - if (it == ctx->embd_seq.end()) { - return nullptr; - } - - return it->second.data(); -} - -// llama state API - -// deprecated -size_t llama_get_state_size(struct llama_context * ctx) { - return llama_state_get_size(ctx); -} - -// deprecated -size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) { - return llama_state_get_data(ctx, dst, -1); -} - -// deprecated -size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) { - return llama_state_set_data(ctx, src, -1); -} - -// deprecated -bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { - return llama_state_load_file(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out); -} - -// deprecated -bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) { - return llama_state_save_file(ctx, path_session, tokens, n_token_count); -} - -// TODO: replace all non-fatal assertions with returned errors or exceptions -struct llama_data_write { - virtual void write(const void * src, size_t size) = 0; - virtual void write_tensor_data(const struct ggml_tensor * tensor, size_t offset, size_t size) = 0; - virtual size_t get_size_written() = 0; - virtual ~llama_data_write() = default; - - void write_string(const std::string & str) { - uint32_t str_size = str.size(); - - write(&str_size, sizeof(str_size)); - write(str.data(), str_size); - } - - void write_model_info(const struct llama_context * ctx) { - const std::string arch_str = llm_arch_name(ctx->model.arch); - write_string(arch_str); - // TODO: add more model-specific info which should prevent loading the session file if not identical - } - - //void write_rng(const std::mt19937 & rng) { - // std::ostringstream rng_ss; - // rng_ss << rng; - - // const std::string & rng_str = rng_ss.str(); - - // write_string(rng_str); - //} - - void write_output_ids(struct llama_context * ctx) { - llama_output_reorder(*ctx); - - const uint32_t n_outputs = ctx->n_outputs; - - std::vector output_pos; - - const size_t n_batch = ctx->cparams.n_batch; - const auto & output_ids = ctx->output_ids; - - GGML_ASSERT(n_outputs <= ctx->output_size); - - output_pos.resize(n_outputs); - - // build a more compact representation of the output ids - for (size_t i = 0; i < n_batch; ++i) { - // map an output id to a position in the batch - int32_t pos = output_ids[i]; - if (pos >= 0) { - GGML_ASSERT((uint32_t) pos < n_outputs); - output_pos[pos] = i; + if (!cparams.offload_kqv) { + if (strcmp(name, "kqv_merged_cont") == 0) { + // all nodes between the KV store and the attention output are run on the CPU + ggml_backend_sched_set_tensor_backend(sched.get(), cur, backend_cpu); } } - write(&n_outputs, sizeof(n_outputs)); - - if (n_outputs) { - write(output_pos.data(), n_outputs * sizeof(int32_t)); - } - } - - void write_logits(const struct llama_context * ctx) { - const uint64_t logits_size = std::min((uint64_t) ctx->logits_size, (uint64_t) ctx->n_outputs * ctx->model.vocab.n_tokens()); - - write(&logits_size, sizeof(logits_size)); - - if (logits_size) { - write(ctx->logits, logits_size * sizeof(float)); - } - } - - void write_embeddings(const struct llama_context * ctx) { - const uint64_t embeddings_size = std::min((uint64_t) ctx->embd_size, (uint64_t) ctx->n_outputs * ctx->model.hparams.n_embd); - - write(&embeddings_size, sizeof(embeddings_size)); - - if (embeddings_size) { - write(ctx->embd, embeddings_size * sizeof(float)); - } - } - - void write_kv_cache_meta(const llama_kv_cache & kv_self, const std::vector> & cell_ranges, llama_seq_id seq_id = -1) { - for (const auto & range : cell_ranges) { - for (uint32_t i = range.first; i < range.second; ++i) { - const auto & cell = kv_self.cells[i]; - const llama_pos pos = cell.pos; - const uint32_t n_seq_id = seq_id == -1 ? cell.seq_id.size() : 0; - - write(&pos, sizeof(pos)); - write(&n_seq_id, sizeof(n_seq_id)); - - if (n_seq_id) { - for (auto seq_id : cell.seq_id) { - write(&seq_id, sizeof(seq_id)); - } - } - } - } - } - - void write_kv_cache_data(const struct llama_context * ctx, const std::vector> & cell_ranges) { - const struct llama_kv_cache & kv_self = ctx->kv_self; - const struct llama_hparams & hparams = ctx->model.hparams; - - const uint32_t v_trans = kv_self.v_trans ? 1 : 0; - const uint32_t n_layer = hparams.n_layer; - - write(&v_trans, sizeof(v_trans)); - write(&n_layer, sizeof(n_layer)); - - std::vector tmp_buf; - - // Iterate and write all the keys first, each row is a cell - // Get whole range at a time - for (uint32_t il = 0; il < n_layer; ++il) { - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); - - // Write key type - const int32_t k_type_i = (int32_t)kv_self.k_l[il]->type; - write(&k_type_i, sizeof(k_type_i)); - - // Write row size of key - const uint64_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa); - write(&k_size_row, sizeof(k_size_row)); - - // Read each range of cells of k_size length each into tmp_buf and write out - for (const auto & range : cell_ranges) { - const size_t range_size = range.second - range.first; - const size_t buf_size = range_size * k_size_row; - write_tensor_data(kv_self.k_l[il], range.first * k_size_row, buf_size); - } - } - - if (!kv_self.v_trans) { - for (uint32_t il = 0; il < n_layer; ++il) { - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); - - // Write value type - const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; - write(&v_type_i, sizeof(v_type_i)); - - // Write row size of value - const uint64_t v_size_row = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa); - write(&v_size_row, sizeof(v_size_row)); - - // Read each range of cells of v_size length each into tmp_buf and write out - for (const auto & range : cell_ranges) { - const size_t range_size = range.second - range.first; - const size_t buf_size = range_size * v_size_row; - write_tensor_data(kv_self.v_l[il], range.first * v_size_row, buf_size); - } - } - } else { - // When v is transposed, we also need the element size and get the element ranges from each row - const uint32_t kv_size = kv_self.size; - for (uint32_t il = 0; il < n_layer; ++il) { - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); - - // Write value type - const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; - write(&v_type_i, sizeof(v_type_i)); - - // Write element size - const uint32_t v_size_el = ggml_type_size(kv_self.v_l[il]->type); - write(&v_size_el, sizeof(v_size_el)); - - // Write GQA embedding size - write(&n_embd_v_gqa, sizeof(n_embd_v_gqa)); - - // For each row, we get the element values of each cell - for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { - // Read each range of cells of v_size_el length each into tmp_buf and write out - for (const auto & range : cell_ranges) { - const size_t range_size = range.second - range.first; - const size_t src_offset = (range.first + j * kv_size) * v_size_el; - const size_t buf_size = range_size * v_size_el; - write_tensor_data(kv_self.v_l[il], src_offset, buf_size); - } - } - } - } - } - - void write_kv_cache(const struct llama_context * ctx, llama_seq_id seq_id = -1) { - const struct llama_kv_cache & kv_self = ctx->kv_self; - std::vector> cell_ranges; // ranges, from inclusive, to exclusive - uint32_t cell_count = 0; - - // Count the number of cells with the specified seq_id - // Find all the ranges of cells with this seq id (or all, when -1) - uint32_t cell_range_begin = kv_self.size; - for (uint32_t i = 0; i < kv_self.size; ++i) { - const auto & cell = kv_self.cells[i]; - if ((seq_id == -1 && !cell.is_empty()) || cell.has_seq_id(seq_id)) { - ++cell_count; - if (cell_range_begin == kv_self.size) { - cell_range_begin = i; - } - } else { - if (cell_range_begin != kv_self.size) { - cell_ranges.emplace_back(cell_range_begin, i); - cell_range_begin = kv_self.size; - } - } - } - if (cell_range_begin != kv_self.size) { - cell_ranges.emplace_back(cell_range_begin, kv_self.size); - } - - // DEBUG CHECK: Sum of cell counts in ranges should equal the total cell count - uint32_t cell_count_check = 0; - for (const auto & range : cell_ranges) { - cell_count_check += range.second - range.first; - } - GGML_ASSERT(cell_count == cell_count_check); - - write(&cell_count, sizeof(cell_count)); - - write_kv_cache_meta(kv_self, cell_ranges, seq_id); - write_kv_cache_data(ctx, cell_ranges); - } -}; - -struct llama_data_read { - virtual const uint8_t * read(size_t size) = 0; - virtual void read_to(void * dst, size_t size) = 0; - virtual size_t get_size_read() = 0; - virtual ~llama_data_read() = default; - - void read_string(std::string & str) { - uint32_t str_size; - read_to(&str_size, sizeof(str_size)); - - str.assign((const char *) read(str_size), str_size); - } - - // validate model information - void read_model_info(const struct llama_context * ctx) { - const std::string cur_arch_str = llm_arch_name(ctx->model.arch); - - std::string arch_str; - read_string(arch_str); - if (cur_arch_str != arch_str) { - throw std::runtime_error(format("wrong model arch: '%s' instead of '%s'", arch_str.c_str(), cur_arch_str.c_str())); - } - // TODO: add more info which needs to be identical but which is not verified otherwise - } - - //void read_rng(std::mt19937 & rng) { - // std::string rng_str; - // read_string(rng_str); - - // std::istringstream rng_ss(rng_str); - // rng_ss >> rng; - - // if (rng_ss.fail()) { - // throw std::runtime_error("failed to load RNG state"); - // } - //} - - void read_output_ids(struct llama_context * ctx) { - std::vector output_pos; - - uint32_t n_outputs; - read_to(&n_outputs, sizeof(n_outputs)); - - if (n_outputs > llama_output_reserve(*ctx, n_outputs)) { - throw std::runtime_error("could not reserve outputs"); - } - - if (n_outputs) { - output_pos.resize(n_outputs); - read_to(output_pos.data(), n_outputs * sizeof(int32_t)); - - for (int32_t i = 0; i < (int32_t) output_pos.size(); ++i) { - int32_t id = output_pos[i]; - if ((uint32_t) id >= ctx->cparams.n_batch) { - throw std::runtime_error(format("invalid output id, %d does not fit in batch size of %u", id, ctx->cparams.n_batch)); - } - ctx->output_ids[id] = i; - } - - ctx->n_outputs = n_outputs; - } - } - - void read_logits(struct llama_context * ctx) { - uint64_t logits_size; - read_to(&logits_size, sizeof(logits_size)); - - if (ctx->logits_size < logits_size) { - throw std::runtime_error("logits buffer too small"); - } - - if (logits_size) { - read_to(ctx->logits, logits_size * sizeof(float)); - } - } - - void read_embeddings(struct llama_context * ctx) { - uint64_t embeddings_size; - read_to(&embeddings_size, sizeof(embeddings_size)); - - if (ctx->embd_size < embeddings_size) { - throw std::runtime_error("embeddings buffer too small"); - } - - if (embeddings_size) { - read_to(ctx->embd, embeddings_size * sizeof(float)); - } - } - - bool read_kv_cache_meta(struct llama_context * ctx, uint32_t cell_count, llama_seq_id dest_seq_id = -1) { - struct llama_kv_cache & kv_self = ctx->kv_self; - - if (dest_seq_id != -1) { - // single sequence - - llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1); - - llama_ubatch batch = ctx->sbatch.reserve_ubatch(cell_count, /* has_embd */ false); - batch.n_tokens = cell_count; - batch.n_seq_tokens = cell_count; - batch.n_seqs = 1; - - for (uint32_t i = 0; i < cell_count; ++i) { - llama_pos pos; - uint32_t n_seq_id; - - read_to(&pos, sizeof(pos)); - read_to(&n_seq_id, sizeof(n_seq_id)); - - if (n_seq_id != 0) { - LLAMA_LOG_ERROR("%s: invalid seq_id-agnostic kv cell\n", __func__); - return false; - } - - batch.pos[i] = pos; - } - batch.n_seq_id[0] = 1; - batch.seq_id[0] = &dest_seq_id; - if (!llama_kv_cache_find_slot(kv_self, batch)) { - LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__); - return false; - } - - // DEBUG CHECK: kv_self.head should be our first cell, kv_self.head + cell_count - 1 should be our last cell (verify seq_id and pos values) - // Assume that this is one contiguous block of cells - GGML_ASSERT(kv_self.head + cell_count <= kv_self.size); - GGML_ASSERT(kv_self.cells[kv_self.head].pos == batch.pos[0]); - GGML_ASSERT(kv_self.cells[kv_self.head + cell_count - 1].pos == batch.pos[cell_count - 1]); - GGML_ASSERT(kv_self.cells[kv_self.head].has_seq_id(dest_seq_id)); - GGML_ASSERT(kv_self.cells[kv_self.head + cell_count - 1].has_seq_id(dest_seq_id)); - } else { - // whole KV cache restore - - if (cell_count > kv_self.size) { - LLAMA_LOG_ERROR("%s: not enough cells in kv cache\n", __func__); - return false; - } - - llama_kv_cache_clear(kv_self); - - for (uint32_t i = 0; i < cell_count; ++i) { - llama_kv_cell & cell = kv_self.cells[i]; - - llama_pos pos; - uint32_t n_seq_id; - - read_to(&pos, sizeof(pos)); - read_to(&n_seq_id, sizeof(n_seq_id)); - - cell.pos = pos; - - for (uint32_t j = 0; j < n_seq_id; ++j) { - llama_seq_id seq_id; - read_to(&seq_id, sizeof(seq_id)); - - if (seq_id < 0 || (uint32_t) seq_id >= llama_n_seq_max(ctx)) { - LLAMA_LOG_ERROR("%s: invalid seq_id, %d is out of range [0, %u)\n", __func__, seq_id, llama_n_seq_max(ctx)); - return false; - } - - cell.seq_id.insert(seq_id); - - if (kv_self.recurrent) { - int32_t & tail = kv_self.cells[seq_id].tail; - if (tail != -1) { - LLAMA_LOG_ERROR("%s: duplicate tail for seq_id %d in cell %d and %d\n", __func__, seq_id, i, tail); - return false; + // norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends + // FIXME: fix in ggml_backend_sched + const bool full_offload = model.params.n_gpu_layers > (int) model.hparams.n_layer; + if (ubatch.n_tokens < 32 || full_offload) { + if (il != -1 && strcmp(name, "norm") == 0) { + const auto & dev_layer = model.dev_layer(il); + for (const auto & backend : backends) { + if (ggml_backend_get_device(backend.get()) == dev_layer) { + if (ggml_backend_supports_op(backend.get(), cur)) { + ggml_backend_sched_set_tensor_backend(sched.get(), cur, backend.get()); } - tail = i; - } - } - } - - kv_self.head = 0; - kv_self.used = cell_count; - } - - if (kv_self.recurrent) { - for (uint32_t i = 0; i < cell_count; ++i) { - uint32_t cell_id = kv_self.head + i; - // make sure the recurrent states will keep their restored state - kv_self.cells[cell_id].src = cell_id; - } - } - - return true; - } - - bool read_kv_cache_data(struct llama_context * ctx, uint32_t cell_count) { - const struct llama_hparams & hparams = ctx->model.hparams; - struct llama_kv_cache & kv_self = ctx->kv_self; - uint32_t v_trans; - uint32_t n_layer; - read_to(&v_trans, sizeof(v_trans)); - read_to(&n_layer, sizeof(n_layer)); - - if (n_layer != hparams.n_layer) { - LLAMA_LOG_ERROR("%s: mismatched layer count (%u instead of %u)\n", __func__, n_layer, hparams.n_layer); - return false; - } - if (cell_count > kv_self.size) { - LLAMA_LOG_ERROR("%s: not enough cells in kv cache to restore state (%u > %u)\n", __func__, cell_count, kv_self.size); - return false; - } - if (kv_self.v_trans != (bool) v_trans) { - LLAMA_LOG_ERROR("%s: incompatible V transposition\n", __func__); - return false; - } - - // For each layer, read the keys for each cell, one row is one cell, read as one contiguous block - for (uint32_t il = 0; il < n_layer; ++il) { - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); - - // Read type of key - int32_t k_type_i_ref; - read_to(&k_type_i_ref, sizeof(k_type_i_ref)); - const int32_t k_type_i = (int32_t)kv_self.k_l[il]->type; - if (k_type_i != k_type_i_ref) { - LLAMA_LOG_ERROR("%s: mismatched key type (%d != %d, layer %d)\n", __func__, k_type_i, k_type_i_ref, il); - return false; - } - - // Read row size of key - uint64_t k_size_row_ref; - read_to(&k_size_row_ref, sizeof(k_size_row_ref)); - const size_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa); - if (k_size_row != k_size_row_ref) { - LLAMA_LOG_ERROR("%s: mismatched key row size (%zu != %zu, layer %d)\n", __func__, k_size_row, (size_t) k_size_row_ref, il); - return false; - } - - if (cell_count) { - // Read and set the keys for the whole cell range - ggml_backend_tensor_set(kv_self.k_l[il], read(cell_count * k_size_row), kv_self.head * k_size_row, cell_count * k_size_row); - } - } - - if (!kv_self.v_trans) { - for (uint32_t il = 0; il < n_layer; ++il) { - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); - - // Read type of value - int32_t v_type_i_ref; - read_to(&v_type_i_ref, sizeof(v_type_i_ref)); - const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; - if (v_type_i != v_type_i_ref) { - LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il); - return false; - } - - // Read row size of value - uint64_t v_size_row_ref; - read_to(&v_size_row_ref, sizeof(v_size_row_ref)); - const size_t v_size_row = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa); - if (v_size_row != v_size_row_ref) { - LLAMA_LOG_ERROR("%s: mismatched value row size (%zu != %zu, layer %d)\n", __func__, v_size_row, (size_t) v_size_row_ref, il); - return false; - } - - if (cell_count) { - // Read and set the values for the whole cell range - ggml_backend_tensor_set(kv_self.v_l[il], read(cell_count * v_size_row), kv_self.head * v_size_row, cell_count * v_size_row); - } - } - } else { - // For each layer, read the values for each cell (transposed) - for (uint32_t il = 0; il < n_layer; ++il) { - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); - - // Read type of value - int32_t v_type_i_ref; - read_to(&v_type_i_ref, sizeof(v_type_i_ref)); - const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; - if (v_type_i != v_type_i_ref) { - LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il); - return false; - } - - // Read element size of value - uint32_t v_size_el_ref; - read_to(&v_size_el_ref, sizeof(v_size_el_ref)); - const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type); - if (v_size_el != v_size_el_ref) { - LLAMA_LOG_ERROR("%s: mismatched value element size (%zu != %zu, layer %d)\n", __func__, v_size_el, (size_t) v_size_el_ref, il); - return false; - } - - // Read GQA embedding size - uint32_t n_embd_v_gqa_ref; - read_to(&n_embd_v_gqa_ref, sizeof(n_embd_v_gqa_ref)); - if (n_embd_v_gqa != n_embd_v_gqa_ref) { - LLAMA_LOG_ERROR("%s: mismatched GQA embedding size (%u != %u, layer %d)\n", __func__, n_embd_v_gqa, n_embd_v_gqa_ref, il); - return false; - } - - if (cell_count) { - // For each row in the transposed matrix, read the values for the whole cell range - for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { - const size_t dst_offset = (kv_self.head + j * kv_self.size) * v_size_el; - ggml_backend_tensor_set(kv_self.v_l[il], read(cell_count * v_size_el), dst_offset, cell_count * v_size_el); } } } } - return true; - } + }; +} - void read_kv_cache(struct llama_context * ctx, llama_seq_id seq_id = -1) { - uint32_t cell_count; - read_to(&cell_count, sizeof(cell_count)); +// +// state save/load +// - bool res = read_kv_cache_meta(ctx, cell_count, seq_id) && read_kv_cache_data(ctx, cell_count); - - if (!res) { - if (seq_id == -1) { - llama_kv_cache_clear(ctx); - } else { - llama_kv_cache_seq_rm(ctx, seq_id, -1, -1); - } - throw std::runtime_error("failed to restore kv cache"); - } - } -}; - -struct llama_data_write_dummy : llama_data_write { - size_t size_written = 0; - - llama_data_write_dummy() {} +class llama_io_write_dummy : public llama_io_write_i { +public: + llama_io_write_dummy() = default; void write(const void * /* src */, size_t size) override { size_written += size; } - void write_tensor_data(const struct ggml_tensor * /* tensor */, size_t /* offset */, size_t size) override { + void write_tensor(const ggml_tensor * /* tensor */, size_t /* offset */, size_t size) override { size_written += size; } - size_t get_size_written() override { + size_t n_bytes() override { return size_written; } + +private: + size_t size_written = 0; }; -struct llama_data_write_buffer : llama_data_write { - uint8_t * ptr; - size_t buf_size = 0; - size_t size_written = 0; - - llama_data_write_buffer(uint8_t * p, size_t len) : ptr(p), buf_size(len) {} +class llama_io_write_buffer : public llama_io_write_i { +public: + llama_io_write_buffer( + uint8_t * p, size_t len) : ptr(p), buf_size(len) {} void write(const void * src, size_t size) override { if (size > buf_size) { @@ -1406,7 +1735,7 @@ struct llama_data_write_buffer : llama_data_write { buf_size -= size; } - void write_tensor_data(const struct ggml_tensor * tensor, size_t offset, size_t size) override { + void write_tensor(const ggml_tensor * tensor, size_t offset, size_t size) override { if (size > buf_size) { throw std::runtime_error("unexpectedly reached end of buffer"); } @@ -1416,17 +1745,19 @@ struct llama_data_write_buffer : llama_data_write { buf_size -= size; } - size_t get_size_written() override { + size_t n_bytes() override { return size_written; } + +private: + uint8_t * ptr; + size_t buf_size = 0; + size_t size_written = 0; }; -struct llama_data_read_buffer : llama_data_read { - const uint8_t * ptr; - size_t buf_size = 0; - size_t size_read = 0; - - llama_data_read_buffer(const uint8_t * p, size_t len) : ptr(p), buf_size(len) {} +class llama_io_read_buffer : public llama_io_read_i { +public: + llama_io_read_buffer(const uint8_t * p, size_t len) : ptr(p), buf_size(len) {} const uint8_t * read(size_t size) override { const uint8_t * base_ptr = ptr; @@ -1443,40 +1774,44 @@ struct llama_data_read_buffer : llama_data_read { memcpy(dst, read(size), size); } - size_t get_size_read() override { + size_t n_bytes() override { return size_read; } + +private: + const uint8_t * ptr; + size_t buf_size = 0; + size_t size_read = 0; }; -struct llama_data_write_file : llama_data_write { - llama_file * file; - size_t size_written = 0; - std::vector temp_buffer; - - llama_data_write_file(llama_file * f) : file(f) {} +class llama_io_write_file : public llama_io_write_i { +public: + llama_io_write_file(llama_file * f) : file(f) {} void write(const void * src, size_t size) override { file->write_raw(src, size); size_written += size; } - void write_tensor_data(const struct ggml_tensor * tensor, size_t offset, size_t size) override { + void write_tensor(const ggml_tensor * tensor, size_t offset, size_t size) override { temp_buffer.resize(size); ggml_backend_tensor_get(tensor, temp_buffer.data(), offset, size); write(temp_buffer.data(), temp_buffer.size()); } - size_t get_size_written() override { + size_t n_bytes() override { return size_written; } + +private: + llama_file * file; + size_t size_written = 0; + std::vector temp_buffer; }; -struct llama_data_read_file : llama_data_read { - llama_file * file; - size_t size_read = 0; - std::vector temp_buffer; - - llama_data_read_file(llama_file * f) : file(f) {} +class llama_io_read_file : public llama_io_read_i { +public: + llama_io_read_file(llama_file * f) : file(f) {} void read_to(void * dst, size_t size) override { file->read_raw(dst, size); @@ -1489,89 +1824,78 @@ struct llama_data_read_file : llama_data_read { return temp_buffer.data(); } - size_t get_size_read() override { + size_t n_bytes() override { return size_read; } + +private: + llama_file * file; + size_t size_read = 0; + std::vector temp_buffer; }; -/** copy state data into either a buffer or file depending on the passed in context - * - * file context: - * llama_file file("/path", "wb"); - * llama_data_write_file data_ctx(&file); - * llama_state_get_data_internal(ctx, data_ctx); - * - * buffer context: - * std::vector buf(max_size, 0); - * llama_data_write_buffer data_ctx(buf.data(), max_size); - * llama_state_get_data_internal(ctx, data_ctx); - * -*/ -static size_t llama_state_get_data_internal(struct llama_context * ctx, llama_data_write & data_ctx) { - llama_synchronize(ctx); - - data_ctx.write_model_info(ctx); - - // copy outputs - data_ctx.write_output_ids(ctx); - data_ctx.write_logits(ctx); - data_ctx.write_embeddings(ctx); - - data_ctx.write_kv_cache(ctx); - - return data_ctx.get_size_written(); -} - -size_t llama_state_get_data(struct llama_context * ctx, uint8_t * dst, size_t size) { - llama_data_write_buffer data_ctx(dst, size); +size_t llama_context::state_get_size() { + llama_io_write_dummy io; try { - return llama_state_get_data_internal(ctx, data_ctx); - } catch (const std::exception & err) { - LLAMA_LOG_ERROR("%s: error saving state: %s\n", __func__, err.what()); - return 0; - } -} - -// Returns the *actual* size of the state. -// Intended to be used when saving to state to a buffer. -size_t llama_state_get_size(struct llama_context * ctx) { - llama_data_write_dummy data_ctx; - try { - return llama_state_get_data_internal(ctx, data_ctx); + return state_write_data(io); } catch (const std::exception & err) { LLAMA_LOG_ERROR("%s: error getting state size: %s\n", __func__, err.what()); return 0; } } -static size_t llama_state_set_data_internal(struct llama_context * ctx, llama_data_read & data_ctx) { - llama_synchronize(ctx); - - data_ctx.read_model_info(ctx); - - // set outputs - data_ctx.read_output_ids(ctx); - data_ctx.read_logits(ctx); - data_ctx.read_embeddings(ctx); - - data_ctx.read_kv_cache(ctx); - - return data_ctx.get_size_read(); +size_t llama_context::state_get_data(uint8_t * dst, size_t size) { + llama_io_write_buffer io(dst, size); + try { + return state_write_data(io); + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: error saving state: %s\n", __func__, err.what()); + return 0; + } } -// Sets the state reading from the specified source address -size_t llama_state_set_data(struct llama_context * ctx, const uint8_t * src, size_t size) { - llama_data_read_buffer data_ctx(src, size); +size_t llama_context::state_set_data(const uint8_t * src, size_t size) { + llama_io_read_buffer io(src, size); try { - return llama_state_set_data_internal(ctx, data_ctx); + return state_read_data(io); } catch (const std::exception & err) { LLAMA_LOG_ERROR("%s: error loading state: %s\n", __func__, err.what()); return 0; } } -static bool llama_state_load_file_internal(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { - llama_file file(path_session, "rb"); +size_t llama_context::state_seq_get_size(llama_seq_id seq_id) { + llama_io_write_dummy io; + try { + return state_seq_write_data(io, seq_id); + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: error getting state size: %s\n", __func__, err.what()); + return 0; + } +} + +size_t llama_context::state_seq_get_data(llama_seq_id seq_id, uint8_t * dst, size_t size) { + llama_io_write_buffer io(dst, size); + try { + return state_seq_write_data(io, seq_id); + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: error saving state: %s\n", __func__, err.what()); + return 0; + } +} + +size_t llama_context::state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size) { + llama_io_read_buffer io(src, size); + try { + return state_seq_read_data(io, seq_id); + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: error loading state: %s\n", __func__, err.what()); + return 0; + } +} + +bool llama_context::state_load_file(const char * filepath, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { + llama_file file(filepath, "rb"); // sanity checks { @@ -1601,28 +1925,20 @@ static bool llama_state_load_file_internal(struct llama_context * ctx, const cha { const size_t n_state_size_cur = file.size() - file.tell(); - llama_data_read_file data_ctx(&file); - const size_t n_read = llama_state_set_data_internal(ctx, data_ctx); + llama_io_read_file io( &file); + const size_t n_read = state_read_data(io); if (n_read != n_state_size_cur) { LLAMA_LOG_ERROR("%s: did not read all of the session file data! size %zu, got %zu\n", __func__, n_state_size_cur, n_read); return false; } } + return true; } -bool llama_state_load_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { - try { - return llama_state_load_file_internal(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out); - } catch (const std::exception & err) { - LLAMA_LOG_ERROR("%s: error loading session file: %s\n", __func__, err.what()); - return false; - } -} - -static bool llama_state_save_file_internal(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) { - llama_file file(path_session, "wb"); +bool llama_context::state_save_file(const char * filepath, const llama_token * tokens, size_t n_token_count) { + llama_file file(filepath, "wb"); file.write_u32(LLAMA_SESSION_MAGIC); file.write_u32(LLAMA_SESSION_VERSION); @@ -1632,82 +1948,13 @@ static bool llama_state_save_file_internal(struct llama_context * ctx, const cha file.write_raw(tokens, sizeof(llama_token) * n_token_count); // save the context state using stream saving - llama_data_write_file data_ctx(&file); - llama_state_get_data_internal(ctx, data_ctx); + llama_io_write_file io(&file); + state_write_data(io); return true; } -bool llama_state_save_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) { - try { - return llama_state_save_file_internal(ctx, path_session, tokens, n_token_count); - } catch (const std::exception & err) { - LLAMA_LOG_ERROR("%s: error saving session file: %s\n", __func__, err.what()); - return false; - } -} - -static size_t llama_state_seq_get_data_internal(struct llama_context * ctx, llama_data_write & data_ctx, llama_seq_id seq_id) { - llama_synchronize(ctx); - - data_ctx.write_kv_cache(ctx, seq_id); - - return data_ctx.get_size_written(); -} - -size_t llama_state_seq_get_size(struct llama_context * ctx, llama_seq_id seq_id) { - llama_data_write_dummy data_ctx; - return llama_state_seq_get_data_internal(ctx, data_ctx, seq_id); -} - -size_t llama_state_seq_get_data(struct llama_context * ctx, uint8_t * dst, size_t size, llama_seq_id seq_id) { - llama_data_write_buffer data_ctx(dst, size); - try { - return llama_state_seq_get_data_internal(ctx, data_ctx, seq_id); - } catch (const std::exception & err) { - LLAMA_LOG_ERROR("%s: error saving sequence state: %s\n", __func__, err.what()); - return 0; - } -} - -static size_t llama_state_seq_set_data_internal(struct llama_context * ctx, llama_data_read & data_ctx, llama_seq_id dest_seq_id) { - llama_synchronize(ctx); - - data_ctx.read_kv_cache(ctx, dest_seq_id); - - return data_ctx.get_size_read(); -} - -size_t llama_state_seq_set_data(struct llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id dest_seq_id) { - llama_data_read_buffer data_ctx(src, size); - try { - return llama_state_seq_set_data_internal(ctx, data_ctx, dest_seq_id); - } catch (const std::exception & err) { - LLAMA_LOG_ERROR("%s: error loading sequence state: %s\n", __func__, err.what()); - return 0; - } -} - -static size_t llama_state_seq_save_file_internal(struct llama_context * ctx, const char * filepath, llama_seq_id seq_id, const llama_token * tokens, size_t n_token_count) { - llama_file file(filepath, "wb"); - - file.write_u32(LLAMA_STATE_SEQ_MAGIC); - file.write_u32(LLAMA_STATE_SEQ_VERSION); - - // save the prompt - file.write_u32((uint32_t) n_token_count); - file.write_raw(tokens, sizeof(llama_token) * n_token_count); - - // save the context state using stream saving - llama_data_write_file data_ctx(&file); - llama_state_seq_get_data_internal(ctx, data_ctx, seq_id); - - const size_t res = file.tell(); - GGML_ASSERT(res == sizeof(uint32_t) * 3 + sizeof(llama_token) * n_token_count + data_ctx.get_size_written()); - return res; -} - -static size_t llama_state_seq_load_file_internal(struct llama_context * ctx, const char * filepath, llama_seq_id dest_seq_id, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { +size_t llama_context::state_seq_load_file(llama_seq_id seq_id, const char * filepath, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { llama_file file(filepath, "rb"); // version checks @@ -1737,8 +1984,8 @@ static size_t llama_state_seq_load_file_internal(struct llama_context * ctx, con // restore the context state { const size_t state_size = file.size() - file.tell(); - llama_data_read_file data_ctx(&file); - const size_t nread = llama_state_seq_set_data_internal(ctx, data_ctx, dest_seq_id); + llama_io_read_file io(&file); + const size_t nread = state_seq_read_data(io, seq_id); if (!nread) { LLAMA_LOG_ERROR("%s: failed to restore sequence state\n", __func__); return 0; @@ -1750,26 +1997,785 @@ static size_t llama_state_seq_load_file_internal(struct llama_context * ctx, con return file.tell(); } -size_t llama_state_seq_save_file(struct llama_context * ctx, const char * filepath, llama_seq_id seq_id, const llama_token * tokens, size_t n_token_count) { +size_t llama_context::state_seq_save_file(llama_seq_id seq_id, const char * filepath, const llama_token * tokens, size_t n_token_count) { + llama_file file(filepath, "wb"); + + file.write_u32(LLAMA_STATE_SEQ_MAGIC); + file.write_u32(LLAMA_STATE_SEQ_VERSION); + + // save the prompt + file.write_u32((uint32_t) n_token_count); + file.write_raw(tokens, sizeof(llama_token) * n_token_count); + + // save the context state using stream saving + llama_io_write_file io(&file); + state_seq_write_data(io, seq_id); + + const size_t res = file.tell(); + GGML_ASSERT(res == sizeof(uint32_t) * 3 + sizeof(llama_token) * n_token_count + io.n_bytes()); + + return res; +} + +size_t llama_context::state_write_data(llama_io_write_i & io) { + LLAMA_LOG_DEBUG("%s: writing state\n", __func__); + + // write model info + { + LLAMA_LOG_DEBUG("%s: - writing model info\n", __func__); + + const std::string arch_str = llm_arch_name(model.arch); + io.write_string(arch_str); + // TODO: add more model-specific info which should prevent loading the session file if not identical + } + + // write output ids + { + LLAMA_LOG_DEBUG("%s: - writing output ids\n", __func__); + + output_reorder(); + + const auto n_outputs = this->n_outputs; + const auto & output_ids = this->output_ids; + + std::vector w_output_pos; + + GGML_ASSERT(n_outputs <= n_outputs_max); + + w_output_pos.resize(n_outputs); + + // build a more compact representation of the output ids + for (size_t i = 0; i < n_batch(); ++i) { + // map an output id to a position in the batch + int32_t pos = output_ids[i]; + if (pos >= 0) { + GGML_ASSERT(pos < n_outputs); + w_output_pos[pos] = i; + } + } + + io.write(&n_outputs, sizeof(n_outputs)); + + if (n_outputs) { + io.write(w_output_pos.data(), n_outputs * sizeof(int32_t)); + } + } + + // write logits + { + LLAMA_LOG_DEBUG("%s: - writing logits\n", __func__); + + const uint64_t logits_size = std::min((uint64_t) this->logits_size, (uint64_t) n_outputs * model.vocab.n_tokens()); + + io.write(&logits_size, sizeof(logits_size)); + + if (logits_size) { + io.write(logits, logits_size * sizeof(float)); + } + } + + // write embeddings + { + LLAMA_LOG_DEBUG("%s: - writing embeddings\n", __func__); + + const uint64_t embd_size = std::min((uint64_t) this->embd_size, (uint64_t) n_outputs * model.hparams.n_embd); + + io.write(&embd_size, sizeof(embd_size)); + + if (embd_size) { + io.write(embd, embd_size * sizeof(float)); + } + } + + LLAMA_LOG_DEBUG("%s: - writing KV self\n", __func__); + kv_self->state_write(io); + + return io.n_bytes(); +} + +size_t llama_context::state_read_data(llama_io_read_i & io) { + LLAMA_LOG_DEBUG("%s: reading state\n", __func__); + + // read model info + { + LLAMA_LOG_DEBUG("%s: - reading model info\n", __func__); + + const std::string cur_arch_str = llm_arch_name(model.arch); + + std::string arch_str; + io.read_string(arch_str); + if (cur_arch_str != arch_str) { + throw std::runtime_error(format("wrong model arch: '%s' instead of '%s'", arch_str.c_str(), cur_arch_str.c_str())); + } + // TODO: add more info which needs to be identical but which is not verified otherwise + } + + // read output ids + { + LLAMA_LOG_DEBUG("%s: - reading output ids\n", __func__); + + auto n_outputs = this->n_outputs; + io.read_to(&n_outputs, sizeof(n_outputs)); + + if (n_outputs > output_reserve(n_outputs)) { + throw std::runtime_error("could not reserve outputs"); + } + + std::vector output_pos; + + if (n_outputs) { + output_pos.resize(n_outputs); + io.read_to(output_pos.data(), n_outputs * sizeof(int32_t)); + + for (int32_t i = 0; i < (int32_t) output_pos.size(); ++i) { + int32_t id = output_pos[i]; + if ((uint32_t) id >= n_batch()) { + throw std::runtime_error(format("invalid output id, %d does not fit in batch size of %u", id, n_batch())); + } + this->output_ids[id] = i; + } + + this->n_outputs = n_outputs; + } + } + + // read logits + { + LLAMA_LOG_DEBUG("%s: - reading logits\n", __func__); + + uint64_t logits_size; + io.read_to(&logits_size, sizeof(logits_size)); + + if (this->logits_size < logits_size) { + throw std::runtime_error("logits buffer too small"); + } + + if (logits_size) { + io.read_to(this->logits, logits_size * sizeof(float)); + } + } + + // read embeddings + { + LLAMA_LOG_DEBUG("%s: - reading embeddings\n", __func__); + + uint64_t embd_size; + io.read_to(&embd_size, sizeof(embd_size)); + + if (this->embd_size < embd_size) { + throw std::runtime_error("embeddings buffer too small"); + } + + if (embd_size) { + io.read_to(this->embd, embd_size * sizeof(float)); + } + } + + LLAMA_LOG_DEBUG("%s: - reading KV self\n", __func__); + kv_self->state_read(io); + + return io.n_bytes(); +} + +size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id) { + GGML_UNUSED(seq_id); + + kv_self->state_write(io, seq_id); + + return io.n_bytes(); +} + +size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id) { + GGML_UNUSED(seq_id); + + kv_self->state_read(io, seq_id); + + return io.n_bytes(); +} + +// +// perf +// + +llama_perf_context_data llama_context::perf_get_data() const { + llama_perf_context_data data = {}; + + data.t_start_ms = 1e-3 * t_start_us; + data.t_load_ms = 1e-3 * t_load_us; + data.t_p_eval_ms = 1e-3 * t_p_eval_us; + data.t_eval_ms = 1e-3 * t_eval_us; + data.n_p_eval = std::max(1, n_p_eval); + data.n_eval = std::max(1, n_eval); + + return data; +} + +void llama_context::perf_reset() { + t_start_us = ggml_time_us(); + t_eval_us = n_eval = 0; + t_p_eval_us = n_p_eval = 0; +} + +// +// interface implementation +// + +llama_context_params llama_context_default_params() { + llama_context_params result = { + /*.n_ctx =*/ 512, + /*.n_batch =*/ 2048, + /*.n_ubatch =*/ 512, + /*.n_seq_max =*/ 1, + /*.n_threads =*/ GGML_DEFAULT_N_THREADS, // TODO: better default + /*.n_threads_batch =*/ GGML_DEFAULT_N_THREADS, + /*.rope_scaling_type =*/ LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED, + /*.pooling_type =*/ LLAMA_POOLING_TYPE_UNSPECIFIED, + /*.attention_type =*/ LLAMA_ATTENTION_TYPE_UNSPECIFIED, + /*.rope_freq_base =*/ 0.0f, + /*.rope_freq_scale =*/ 0.0f, + /*.yarn_ext_factor =*/ -1.0f, + /*.yarn_attn_factor =*/ 1.0f, + /*.yarn_beta_fast =*/ 32.0f, + /*.yarn_beta_slow =*/ 1.0f, + /*.yarn_orig_ctx =*/ 0, + /*.defrag_thold =*/ -1.0f, + /*.cb_eval =*/ nullptr, + /*.cb_eval_user_data =*/ nullptr, + /*.type_k =*/ GGML_TYPE_F16, + /*.type_v =*/ GGML_TYPE_F16, + /*.logits_all =*/ false, + /*.embeddings =*/ false, + /*.offload_kqv =*/ true, + /*.flash_attn =*/ false, + /*.no_perf =*/ true, + /*.abort_callback =*/ nullptr, + /*.abort_callback_data =*/ nullptr, + }; + + return result; +} + +llama_context * llama_init_from_model( + llama_model * model, + llama_context_params params) { + if (!model) { + LLAMA_LOG_ERROR("%s: model cannot be NULL\n", __func__); + return nullptr; + } + + if (params.n_batch == 0 && params.n_ubatch == 0) { + LLAMA_LOG_ERROR("%s: n_batch and n_ubatch cannot both be zero\n", __func__); + return nullptr; + } + + if (params.n_ctx == 0 && model->hparams.n_ctx_train == 0) { + LLAMA_LOG_ERROR("%s: n_ctx and model->hparams.n_ctx_train cannot both be zero\n", __func__); + return nullptr; + } + + if (params.flash_attn && model->arch == LLM_ARCH_GROK) { + LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__); + params.flash_attn = false; + } + + if (params.flash_attn && model->hparams.n_embd_head_k != model->hparams.n_embd_head_v) { + LLAMA_LOG_WARN("%s: flash_attn requires n_embd_head_k == n_embd_head_v - forcing off\n", __func__); + params.flash_attn = false; + } + + if (ggml_is_quantized(params.type_v) && !params.flash_attn) { + LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__); + return nullptr; + } + try { - return llama_state_seq_save_file_internal(ctx, filepath, seq_id, tokens, n_token_count); + auto * ctx = new llama_context(*model, params); + return ctx; + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: failed to initialize the context: %s\n", __func__, err.what()); + } + + return nullptr; +} + +// deprecated +llama_context * llama_new_context_with_model( + llama_model * model, + llama_context_params params) { + return llama_init_from_model(model, params); +} + +void llama_free(llama_context * ctx) { + delete ctx; +} + +uint32_t llama_n_ctx(const llama_context * ctx) { + return ctx->n_ctx(); +} + +uint32_t llama_n_batch(const llama_context * ctx) { + return ctx->n_batch(); +} + +uint32_t llama_n_ubatch(const llama_context * ctx) { + return ctx->n_ubatch(); +} + +uint32_t llama_n_seq_max(const llama_context * ctx) { + return ctx->n_seq_max(); +} + +const llama_model * llama_get_model(const llama_context * ctx) { + return &ctx->get_model(); +} + +llama_kv_cache * llama_get_kv_self(llama_context * ctx) { + return ctx->get_kv_self(); +} + +void llama_kv_self_update(llama_context * ctx) { + ctx->kv_self_update(); +} + +enum llama_pooling_type llama_pooling_type(const llama_context * ctx) { + return ctx->pooling_type(); +} + +void llama_attach_threadpool( + llama_context * ctx, + ggml_threadpool_t threadpool, + ggml_threadpool_t threadpool_batch) { + ctx->attach_threadpool(threadpool, threadpool_batch); +} + +void llama_detach_threadpool(llama_context * ctx) { + ctx->detach_threadpool(); +} + +void llama_set_n_threads(llama_context * ctx, int32_t n_threads, int32_t n_threads_batch) { + ctx->set_n_threads(n_threads, n_threads_batch); +} + +int32_t llama_n_threads(llama_context * ctx) { + return ctx->n_threads(); +} + +int32_t llama_n_threads_batch(llama_context * ctx) { + return ctx->n_threads_batch(); +} + +void llama_set_abort_callback(llama_context * ctx, bool (*abort_callback)(void * data), void * abort_callback_data) { + ctx->set_abort_callback(abort_callback, abort_callback_data); +} + +void llama_set_embeddings(llama_context * ctx, bool embeddings) { + ctx->set_embeddings(embeddings); +} + +void llama_set_causal_attn(llama_context * ctx, bool causal_attn) { + ctx->set_causal_attn(causal_attn); +} + +void llama_synchronize(llama_context * ctx) { + ctx->synchronize(); +} + +float * llama_get_logits(llama_context * ctx) { + ctx->synchronize(); + + return ctx->get_logits(); +} + +float * llama_get_logits_ith(llama_context * ctx, int32_t i) { + ctx->synchronize(); + + return ctx->get_logits_ith(i); +} + +float * llama_get_embeddings(llama_context * ctx) { + ctx->synchronize(); + + return ctx->get_embeddings(); +} + +float * llama_get_embeddings_ith(llama_context * ctx, int32_t i) { + ctx->synchronize(); + + return ctx->get_embeddings_ith(i); +} + +float * llama_get_embeddings_seq(llama_context * ctx, llama_seq_id seq_id) { + ctx->synchronize(); + + return ctx->get_embeddings_seq(seq_id); +} + +// llama adapter API + +int32_t llama_set_adapter_lora( + llama_context * ctx, + llama_adapter_lora * adapter, + float scale) { + ctx->set_adapter_lora(adapter, scale); + + return 0; +} + +int32_t llama_rm_adapter_lora( + llama_context * ctx, + llama_adapter_lora * adapter) { + bool res = ctx->rm_adapter_lora(adapter); + + return res ? 0 : -1; +} + +void llama_clear_adapter_lora(llama_context * ctx) { + ctx->clear_adapter_lora(); +} + +int32_t llama_apply_adapter_cvec( + llama_context * ctx, + const float * data, + size_t len, + int32_t n_embd, + int32_t il_start, + int32_t il_end) { + bool res = ctx->apply_adapter_cvec(data, len, n_embd, il_start, il_end); + + return res ? 0 : -1; +} + +// +// kv cache view +// + +llama_kv_cache_view llama_kv_cache_view_init(const llama_context * ctx, int32_t n_seq_max) { + const auto * kv = ctx->get_kv_self(); + if (kv == nullptr) { + LLAMA_LOG_WARN("%s: the context does not have a KV cache\n", __func__); + return {}; + } + + return llama_kv_cache_view_init(*kv, n_seq_max); +} + +void llama_kv_cache_view_update(const llama_context * ctx, llama_kv_cache_view * view) { + const auto * kv = ctx->get_kv_self(); + if (kv == nullptr) { + LLAMA_LOG_WARN("%s: the context does not have a KV cache\n", __func__); + return; + } + + llama_kv_cache_view_update(view, kv); +} + +// +// kv cache +// + +// deprecated +int32_t llama_get_kv_cache_token_count(const llama_context * ctx) { + return llama_kv_self_n_tokens(ctx); +} + +int32_t llama_kv_self_n_tokens(const llama_context * ctx) { + return llama_kv_cache_n_tokens(ctx->get_kv_self()); +} + +// deprecated +int32_t llama_get_kv_cache_used_cells(const llama_context * ctx) { + return llama_kv_self_used_cells(ctx); +} + +int32_t llama_kv_self_used_cells(const llama_context * ctx) { + return llama_kv_cache_used_cells(ctx->get_kv_self()); +} + +// deprecated +void llama_kv_cache_clear(llama_context * ctx) { + llama_kv_self_clear(ctx); +} + +void llama_kv_self_clear(llama_context * ctx) { + llama_kv_cache_clear(ctx->get_kv_self()); +} + +// deprecated +bool llama_kv_cache_seq_rm( + llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1) { + return llama_kv_self_seq_rm(ctx, seq_id, p0, p1); +} + +bool llama_kv_self_seq_rm( + llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1) { + return llama_kv_cache_seq_rm(ctx->get_kv_self(), seq_id, p0, p1); +} + +// deprecated +void llama_kv_cache_seq_cp( + llama_context * ctx, + llama_seq_id seq_id_src, + llama_seq_id seq_id_dst, + llama_pos p0, + llama_pos p1) { + return llama_kv_self_seq_cp(ctx, seq_id_src, seq_id_dst, p0, p1); +} + +void llama_kv_self_seq_cp( + llama_context * ctx, + llama_seq_id seq_id_src, + llama_seq_id seq_id_dst, + llama_pos p0, + llama_pos p1) { + return llama_kv_cache_seq_cp(ctx->get_kv_self(), seq_id_src, seq_id_dst, p0, p1); +} + +// deprecated +void llama_kv_cache_seq_keep( + llama_context * ctx, + llama_seq_id seq_id) { + return llama_kv_self_seq_keep(ctx, seq_id); +} + +void llama_kv_self_seq_keep(llama_context * ctx, llama_seq_id seq_id) { + return llama_kv_cache_seq_keep(ctx->get_kv_self(), seq_id); +} + +// deprecated +void llama_kv_cache_seq_add( + llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + llama_pos delta) { + return llama_kv_self_seq_add(ctx, seq_id, p0, p1, delta); +} + +void llama_kv_self_seq_add( + llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + llama_pos delta) { + return llama_kv_cache_seq_add(ctx->get_kv_self(), seq_id, p0, p1, delta); +} + +// deprecated +void llama_kv_cache_seq_div( + llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + int d) { + return llama_kv_self_seq_div(ctx, seq_id, p0, p1, d); +} + +void llama_kv_self_seq_div( + llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + int d) { + return llama_kv_cache_seq_div(ctx->get_kv_self(), seq_id, p0, p1, d); +} + +// deprecated +llama_pos llama_kv_cache_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) { + return llama_kv_self_seq_pos_max(ctx, seq_id); +} + +llama_pos llama_kv_self_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) { + return llama_kv_cache_seq_pos_max(ctx->get_kv_self(), seq_id); +} + +// deprecated +void llama_kv_cache_defrag(llama_context * ctx) { + return llama_kv_self_defrag(ctx); +} + +void llama_kv_self_defrag(llama_context * ctx) { + llama_kv_cache_defrag(ctx->get_kv_self()); +} + +// deprecated +bool llama_kv_cache_can_shift(const llama_context * ctx) { + return llama_kv_self_can_shift(ctx); +} + +bool llama_kv_self_can_shift(const llama_context * ctx) { + return llama_kv_cache_can_shift(ctx->get_kv_self()); +} + +// deprecated +void llama_kv_cache_update(llama_context * ctx) { + llama_kv_self_update(ctx); +} + +// llama state API + +// deprecated +size_t llama_get_state_size(llama_context * ctx) { + return llama_state_get_size(ctx); +} + +// deprecated +size_t llama_copy_state_data(llama_context * ctx, uint8_t * dst) { + return llama_state_get_data(ctx, dst, -1); +} + +// deprecated +size_t llama_set_state_data(llama_context * ctx, const uint8_t * src) { + return llama_state_set_data(ctx, src, -1); +} + +// deprecated +bool llama_load_session_file(llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { + return llama_state_load_file(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out); +} + +// deprecated +bool llama_save_session_file(llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) { + return llama_state_save_file(ctx, path_session, tokens, n_token_count); +} + +// Returns the *actual* size of the state. +// Intended to be used when saving to state to a buffer. +size_t llama_state_get_size(llama_context * ctx) { + return ctx->state_get_size(); +} + +size_t llama_state_get_data(llama_context * ctx, uint8_t * dst, size_t size) { + ctx->synchronize(); + + return ctx->state_get_data(dst, size); +} + +// Sets the state reading from the specified source address +size_t llama_state_set_data(llama_context * ctx, const uint8_t * src, size_t size) { + ctx->synchronize(); + + return ctx->state_set_data(src, size); +} + +bool llama_state_load_file(llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { + ctx->synchronize(); + + try { + return ctx->state_load_file(path_session, tokens_out, n_token_capacity, n_token_count_out); + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: error loading session file: %s\n", __func__, err.what()); + return false; + } +} + +bool llama_state_save_file(llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) { + ctx->synchronize(); + + try { + return ctx->state_save_file(path_session, tokens, n_token_count); + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: error saving session file: %s\n", __func__, err.what()); + return false; + } +} + +size_t llama_state_seq_get_size(llama_context * ctx, llama_seq_id seq_id) { + return ctx->state_seq_get_size(seq_id); +} + +size_t llama_state_seq_get_data(llama_context * ctx, uint8_t * dst, size_t size, llama_seq_id seq_id) { + ctx->synchronize(); + + return ctx->state_seq_get_data(seq_id, dst, size); +} + +size_t llama_state_seq_set_data(llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id seq_id) { + ctx->synchronize(); + + return ctx->state_seq_set_data(seq_id, src, size); +} + +size_t llama_state_seq_save_file(llama_context * ctx, const char * filepath, llama_seq_id seq_id, const llama_token * tokens, size_t n_token_count) { + ctx->synchronize(); + + try { + return ctx->state_seq_save_file(seq_id, filepath, tokens, n_token_count); } catch (const std::exception & err) { LLAMA_LOG_ERROR("%s: error saving sequence state file: %s\n", __func__, err.what()); return 0; } } -size_t llama_state_seq_load_file(struct llama_context * ctx, const char * filepath, llama_seq_id dest_seq_id, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { +size_t llama_state_seq_load_file(llama_context * ctx, const char * filepath, llama_seq_id dest_seq_id, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { + ctx->synchronize(); + try { - return llama_state_seq_load_file_internal(ctx, filepath, dest_seq_id, tokens_out, n_token_capacity, n_token_count_out); + return ctx->state_seq_load_file(dest_seq_id, filepath, tokens_out, n_token_capacity, n_token_count_out); } catch (const std::exception & err) { LLAMA_LOG_ERROR("%s: error loading sequence state file: %s\n", __func__, err.what()); return 0; } } -const std::vector> & llama_internal_get_tensor_map( - struct llama_context * ctx -) { - return ctx->model.tensors_by_name; +/// + +int32_t llama_encode( + llama_context * ctx, + llama_batch batch) { + const int ret = ctx->encode(batch); + if (ret != 0) { + LLAMA_LOG_ERROR("%s: failed to encode, ret = %d\n", __func__, ret); + } + + return ret; +} + +int32_t llama_decode( + llama_context * ctx, + llama_batch batch) { + const int ret = ctx->decode(batch); + if (ret != 0) { + LLAMA_LOG_ERROR("%s: failed to decode, ret = %d\n", __func__, ret); + } + + return ret; +} + +// +// perf +// + +llama_perf_context_data llama_perf_context(const llama_context * ctx) { + llama_perf_context_data data = {}; + + if (ctx == nullptr) { + return data; + } + + data = ctx->perf_get_data(); + + return data; +} + +void llama_perf_context_print(const llama_context * ctx) { + const auto data = llama_perf_context(ctx); + + const double t_end_ms = 1e-3 * ggml_time_us(); + + LLAMA_LOG_INFO("%s: load time = %10.2f ms\n", __func__, data.t_load_ms); + LLAMA_LOG_INFO("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n", + __func__, data.t_p_eval_ms, data.n_p_eval, data.t_p_eval_ms / data.n_p_eval, 1e3 / data.t_p_eval_ms * data.n_p_eval); + LLAMA_LOG_INFO("%s: eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n", + __func__, data.t_eval_ms, data.n_eval, data.t_eval_ms / data.n_eval, 1e3 / data.t_eval_ms * data.n_eval); + LLAMA_LOG_INFO("%s: total time = %10.2f ms / %5d tokens\n", __func__, (t_end_ms - data.t_start_ms), (data.n_p_eval + data.n_eval)); +} + +void llama_perf_context_reset(llama_context * ctx) { + ctx->perf_reset(); } diff --git a/src/llama-context.h b/src/llama-context.h index a9268b2920..88df8950e4 100644 --- a/src/llama-context.h +++ b/src/llama-context.h @@ -3,66 +3,212 @@ #include "llama.h" #include "llama-batch.h" #include "llama-cparams.h" -#include "llama-model.h" -#include "llama-kv-cache.h" +#include "llama-graph.h" #include "llama-adapter.h" #include "ggml-cpp.h" #include -#include #include -#include + +struct llama_model; +struct llama_kv_cache; + +class llama_io_read_i; +class llama_io_write_i; struct llama_context { - llama_context(const llama_model & model) - : model(model) - , t_start_us(model.t_start_us) - , t_load_us(model.t_load_us) {} + // init scheduler and compute buffers, reserve worst-case graphs + llama_context( + const llama_model & model, + llama_context_params params); - const struct llama_model & model; + ~llama_context(); - struct llama_cparams cparams; - struct llama_sbatch sbatch; // TODO: revisit if needed - struct llama_kv_cache kv_self; - struct llama_adapter_cvec cvec; + void synchronize(); - std::unordered_map lora; + const llama_model & get_model() const; - std::vector backends; - std::vector> set_n_threads_fns; + uint32_t n_ctx() const; + uint32_t n_ctx_per_seq() const; + uint32_t n_batch() const; + uint32_t n_ubatch() const; + uint32_t n_seq_max() const; - ggml_backend_t backend_cpu = nullptr; + uint32_t n_threads() const; + uint32_t n_threads_batch() const; - ggml_threadpool_t threadpool = nullptr; - ggml_threadpool_t threadpool_batch = nullptr; + llama_kv_cache * get_kv_self(); + const llama_kv_cache * get_kv_self() const; - bool has_evaluated_once = false; + void kv_self_update(); - mutable int64_t t_start_us; - mutable int64_t t_load_us; - mutable int64_t t_p_eval_us = 0; - mutable int64_t t_eval_us = 0; + enum llama_pooling_type pooling_type() const; - mutable int64_t t_compute_start_us = 0; - mutable int64_t n_queued_tokens = 0; + float * get_logits(); + float * get_logits_ith(int32_t i); - mutable int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1) - mutable int32_t n_eval = 0; // number of eval calls + float * get_embeddings(); + float * get_embeddings_ith(int32_t i); + float * get_embeddings_seq(llama_seq_id seq_id); - // host buffer for the model output (logits and embeddings) - ggml_backend_buffer_ptr buf_output; + void attach_threadpool( + ggml_threadpool_t threadpool, + ggml_threadpool_t threadpool_batch); + + void detach_threadpool(); + + void set_n_threads(int32_t n_threads, int32_t n_threads_batch); + + void set_abort_callback(bool (*abort_callback)(void * data), void * abort_callback_data); + + void set_embeddings (bool value); + void set_causal_attn(bool value); + + void set_adapter_lora( + llama_adapter_lora * adapter, + float scale); + + bool rm_adapter_lora( + llama_adapter_lora * adapter); + + void clear_adapter_lora(); + + bool apply_adapter_cvec( + const float * data, + size_t len, + int32_t n_embd, + int32_t il_start, + int32_t il_end); + + int encode(llama_batch & inp_batch); + int decode(llama_batch & inp_batch); + + // + // state save/load + // + + size_t state_get_size(); + size_t state_get_data( uint8_t * dst, size_t size); + size_t state_set_data(const uint8_t * src, size_t size); + + size_t state_seq_get_size(llama_seq_id seq_id); + size_t state_seq_get_data(llama_seq_id seq_id, uint8_t * dst, size_t size); + size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size); + + bool state_load_file( + const char * filepath, + llama_token * tokens_out, + size_t n_token_capacity, + size_t * n_token_count_out); + + bool state_save_file( + const char * filepath, + const llama_token * tokens, + size_t n_token_count); + + size_t state_seq_load_file( + llama_seq_id seq_id, + const char * filepath, + llama_token * tokens_out, + size_t n_token_capacity, + size_t * n_token_count_out); + + size_t state_seq_save_file( + llama_seq_id seq_id, + const char * filepath, + const llama_token * tokens, + size_t n_token_count); + + // + // perf + // + + llama_perf_context_data perf_get_data() const; + void perf_reset(); + +private: + // + // output + // + + // Make sure enough space is available for outputs. + // Returns max number of outputs for which space was reserved. + int32_t output_reserve(int32_t n_outputs); + + // make the outputs have the same order they had in the user-provided batch + // TODO: maybe remove this + void output_reorder(); + + // + // graph + // + + int32_t graph_max_nodes() const; + + // zero-out inputs and create the ctx_compute for the compute graph + ggml_cgraph * graph_init(); + + llm_graph_result_ptr graph_build( + ggml_context * ctx, + ggml_cgraph * gf, + const llama_ubatch & ubatch, + llm_graph_type gtype); + + // returns the result of ggml_backend_sched_graph_compute_async execution + ggml_status graph_compute( + ggml_cgraph * gf, + bool batched); + + llm_graph_cb graph_get_cb() const; + + // used by kv_self_update() + ggml_tensor * build_rope_shift( + ggml_context * ctx0, + ggml_tensor * cur, + ggml_tensor * shift, + ggml_tensor * factors, + float freq_base, + float freq_scale, + ggml_backend_buffer * bbuf) const; + + llm_graph_result_ptr build_kv_self_shift( + ggml_context * ctx0, + ggml_cgraph * gf) const; + + llm_graph_result_ptr build_kv_self_defrag( + ggml_context * ctx0, + ggml_cgraph * gf) const; + + // TODO: read/write lora adapters and cvec + size_t state_write_data(llama_io_write_i & io); + size_t state_read_data (llama_io_read_i & io); + + size_t state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id); + size_t state_seq_read_data (llama_io_read_i & io, llama_seq_id seq_id); + + // + // members + // + + const llama_model & model; + + llama_cparams cparams; + llama_adapter_cvec cvec; + llama_adapter_loras loras; + llama_sbatch sbatch; + + llama_cross cross; // TODO: tmp for handling cross-attention - need something better probably + + std::unique_ptr kv_self; + + // TODO: remove + bool logits_all = false; // decode output (2-dimensional array: [n_outputs][n_vocab]) size_t logits_size = 0; // capacity (of floats) for logits float * logits = nullptr; - std::vector output_ids; // map batch token positions to ids of the logits and embd buffers - size_t output_size = 0; // capacity (of tokens positions) for the output buffers - int32_t n_outputs = 0; // number of actually-used outputs in the current ubatch or last logical batch - - bool logits_all = false; - // embeddings output (2-dimensional array: [n_outputs][n_embd]) // populated only when pooling_type == LLAMA_POOLING_TYPE_NONE size_t embd_size = 0; // capacity (of floats) for embeddings @@ -72,57 +218,47 @@ struct llama_context { // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE std::map> embd_seq; - // whether we are computing encoder output or decoder output - bool is_encoding = false; + int32_t n_outputs = 0; // number of actually-used outputs in the current ubatch or last logical batch + int32_t n_outputs_max = 0; // capacity (of tokens positions) for the output buffers - // TODO: find a better way to accommodate mutli-dimension position encoding methods - // number of position id each token get, 1 for each token in most cases. - // when using m-rope, it will be 3 position ids per token to representing 3 dimension coordinate. - int n_pos_per_token = 1; + std::vector output_ids; // map batch token positions to ids of the logits and embd buffers - // output of the encoder part of the encoder-decoder models - std::vector embd_enc; - std::vector> seq_ids_enc; - - // memory buffers used to evaluate the model - std::vector buf_compute_meta; ggml_backend_sched_ptr sched; + ggml_backend_t backend_cpu = nullptr; + std::vector backends; + + ggml_context_ptr ctx_compute; + + ggml_threadpool_t threadpool = nullptr; + ggml_threadpool_t threadpool_batch = nullptr; + ggml_abort_callback abort_callback = nullptr; void * abort_callback_data = nullptr; - // input tensors - struct ggml_tensor * inp_tokens; // I32 [n_batch] - struct ggml_tensor * inp_embd; // F32 [n_embd, n_batch] - struct ggml_tensor * inp_pos; // I32 [n_batch] - struct ggml_tensor * inp_out_ids; // I32 [n_outputs] - struct ggml_tensor * inp_KQ_mask; // F32 [kv_size, n_batch] - struct ggml_tensor * inp_KQ_mask_swa; // F32 [kv_size, n_batch] - struct ggml_tensor * inp_K_shift; // I32 [kv_size] - struct ggml_tensor * inp_mean; // F32 [n_batch, n_batch] - struct ggml_tensor * inp_cls; // I32 [n_batch] - struct ggml_tensor * inp_s_copy; // I32 [kv_size] - struct ggml_tensor * inp_s_mask; // F32 [1, n_kv] - struct ggml_tensor * inp_s_seq; // I32 [n_kv, n_batch] - struct ggml_tensor * inp_pos_bucket; // I32 [n_batch|n_kv, n_batch] - struct ggml_tensor * inp_embd_enc; // F32 [n_embd, n_outputs_enc] - struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch] + std::vector> set_n_threads_fns; + + // buffer types used for the compute buffer of each backend + std::vector backend_ptrs; + std::vector backend_buft; + + // memory buffers used to evaluate the model + std::vector buf_compute_meta; + + // host buffer for the model output (logits and embeddings) + ggml_backend_buffer_ptr buf_output; + + bool has_evaluated_once = false; + + // perf + mutable int64_t t_start_us = 0; + mutable int64_t t_load_us = 0; + mutable int64_t t_p_eval_us = 0; + mutable int64_t t_eval_us = 0; + + mutable int64_t t_compute_start_us = 0; + mutable int64_t n_queued_tokens = 0; + + mutable int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1) + mutable int32_t n_eval = 0; // number of eval calls }; - -// TODO: make these methods of llama_context -void llama_set_k_shift(struct llama_context & lctx); - -void llama_set_s_copy(struct llama_context & lctx); - -void llama_set_inputs(llama_context & lctx, const llama_ubatch & ubatch); - -// Make sure enough space is available for outputs. -// Returns max number of outputs for which space was reserved. -size_t llama_output_reserve(struct llama_context & lctx, size_t n_outputs); - -// make the outputs have the same order they had in the user-provided batch -void llama_output_reorder(struct llama_context & ctx); - -// For internal test use -// TODO: remove -const std::vector> & llama_internal_get_tensor_map(struct llama_context * ctx); diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp new file mode 100644 index 0000000000..4a53e83929 --- /dev/null +++ b/src/llama-graph.cpp @@ -0,0 +1,1668 @@ +#include "llama-graph.h" + +#include "llama-impl.h" +#include "llama-batch.h" +#include "llama-cparams.h" +#include "llama-kv-cache.h" + +#include +#include +#include + +static int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buckets, bool bidirectional) { + // TODO move to hparams if a T5 variant appears that uses a different value + const int64_t max_distance = 128; + + if (bidirectional) { + n_buckets >>= 1; + } + + const int64_t max_exact = n_buckets >> 1; + + int32_t relative_position = x - y; + int32_t relative_bucket = 0; + + if (bidirectional) { + relative_bucket += (relative_position > 0) * n_buckets; + relative_position = abs(relative_position); + } else { + relative_position = -std::min(relative_position, 0); + } + + int32_t relative_position_if_large = floorf(max_exact + logf(1.0 * relative_position / max_exact) * (n_buckets - max_exact) / log(1.0 * max_distance / max_exact)); + relative_position_if_large = std::min(relative_position_if_large, n_buckets - 1); + relative_bucket += (relative_position < max_exact ? relative_position : relative_position_if_large); + + return relative_bucket; +} + +void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) { + if (ubatch->token) { + const int64_t n_tokens = ubatch->n_tokens; + + ggml_backend_tensor_set(tokens, ubatch->token, 0, n_tokens*ggml_element_size(tokens)); + } + + if (ubatch->embd) { + const int64_t n_embd = embd->ne[0]; + const int64_t n_tokens = ubatch->n_tokens; + + ggml_backend_tensor_set(embd, ubatch->embd, 0, n_tokens*n_embd*ggml_element_size(embd)); + } +} + +void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) { + if (ubatch->pos && pos) { + const int64_t n_tokens = ubatch->n_tokens; + + ggml_backend_tensor_set(pos, ubatch->pos, 0, n_tokens*n_pos_per_token*ggml_element_size(pos)); + } +} + +void llm_graph_input_pos_bucket::set_input(const llama_ubatch * ubatch) { + if (pos_bucket) { + const int64_t n_tokens = ubatch->n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(pos_bucket->buffer)); + GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing + + int32_t * data = (int32_t *) pos_bucket->data; + + for (int h = 0; h < 1; ++h) { + for (int j = 0; j < n_tokens; ++j) { + for (int i = 0; i < n_tokens; ++i) { + data[h*(n_tokens*n_tokens) + j*n_tokens + i] = llama_relative_position_bucket(ubatch->pos[i], ubatch->pos[j], hparams.n_rel_attn_bkts, true); + } + } + } + } +} + +void llm_graph_input_pos_bucket_kv::set_input(const llama_ubatch * ubatch) { + if (pos_bucket) { + const int64_t n_tokens = ubatch->n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(pos_bucket->buffer)); + GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing + + int32_t * data = (int32_t *) pos_bucket->data; + + const int64_t n_kv = kv_self->n; + + for (int h = 0; h < 1; ++h) { + for (int j = 0; j < n_tokens; ++j) { + for (int i = 0; i < n_kv; ++i) { + data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(kv_self->cells[i].pos, ubatch->pos[j], hparams.n_rel_attn_bkts, false); + } + } + } + } +} + +void llm_graph_input_out_ids::set_input(const llama_ubatch * ubatch) { + if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) { + //GGML_ASSERT(out_ids && "every model that can must skip unused outputs"); + + if (!out_ids) { + LLAMA_LOG_WARN("%s: 'out_ids' is not created\n", __func__); + } else { + const int64_t n_tokens = ubatch->n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(out_ids->buffer)); + int32_t * data = (int32_t *) out_ids->data; + + if (n_outputs == n_tokens) { + for (int i = 0; i < n_tokens; ++i) { + data[i] = i; + } + } else if (ubatch->output) { + int32_t n_outputs = 0; + for (int i = 0; i < n_tokens; ++i) { + if (ubatch->output[i]) { + data[n_outputs++] = i; + } + } + // the graph needs to have been passed the correct number of outputs + GGML_ASSERT(n_outputs == n_outputs); + } else if (n_outputs == 1) { + // only keep last output + data[0] = n_tokens - 1; + } else { + GGML_ASSERT(n_outputs == 0); + } + } + } +} + +void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) { + if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) { + const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_seq_tokens = ubatch->n_seq_tokens; + const int64_t n_seqs = ubatch->n_seqs; + + GGML_ASSERT(mean); + GGML_ASSERT(ggml_backend_buffer_is_host(mean->buffer)); + + float * data = (float *) mean->data; + memset(mean->data, 0, n_tokens * n_tokens * ggml_element_size(mean)); + + std::vector sum(n_tokens, 0); + + for (int s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[s][0]; + + // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true + GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN"); + + sum[seq_id] += ubatch->n_seq_tokens; + } + + std::vector div(n_tokens, 0.0f); + for (int i = 0; i < n_tokens; ++i) { + const uint64_t s = sum[i]; + if (s > 0) { + div[i] = 1.0f/float(s); + } + } + + for (int s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[s][0]; + + for (int i = 0; i < n_seq_tokens; ++i) { + data[seq_id*n_tokens + s*n_seq_tokens + i] = div[seq_id]; + } + } + } +} + +void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) { + if (cparams.embeddings && ( + cparams.pooling_type == LLAMA_POOLING_TYPE_CLS || + cparams.pooling_type == LLAMA_POOLING_TYPE_RANK)) { + const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_seq_tokens = ubatch->n_seq_tokens; + const int64_t n_seqs = ubatch->n_seqs; + + GGML_ASSERT(cls); + GGML_ASSERT(ggml_backend_buffer_is_host(cls->buffer)); + + uint32_t * data = (uint32_t *) cls->data; + memset(cls->data, 0, n_tokens * ggml_element_size(cls)); + + for (int s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[s][0]; + + // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true + GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS or RANK"); + + for (int i = 0; i < n_seq_tokens; ++i) { + const llama_pos pos = ubatch->pos[s*n_seq_tokens + i]; + + if (pos == 0) { + data[seq_id] = s*n_seq_tokens + i; + } + } + } + } + + if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_LAST) { + const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_seq_tokens = ubatch->n_seq_tokens; + const int64_t n_seqs = ubatch->n_seqs; + + GGML_ASSERT(cls); + GGML_ASSERT(ggml_backend_buffer_is_host(cls->buffer)); + + uint32_t * data = (uint32_t *) cls->data; + memset(cls->data, 0, n_tokens * ggml_element_size(cls)); + + std::vector last_pos(n_tokens, -1); + std::vector last_row(n_tokens, -1); + + for (int s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[s][0]; + + // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true + GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == LAST"); + + for (int i = 0; i < n_seq_tokens; ++i) { + const llama_pos pos = ubatch->pos[s*n_seq_tokens + i]; + + if (pos >= last_pos[seq_id]) { + last_pos[seq_id] = pos; + last_row[seq_id] = s*n_seq_tokens + i; + } + } + } + + for (int i = 0; i < n_tokens; ++i) { + if (last_row[i] >= 0) { + data[i] = last_row[i]; + } + } + } +} + +void llm_graph_input_s_copy::set_input(const llama_ubatch * ubatch) { + GGML_UNUSED(ubatch); + + const int64_t n_kv = kv_self->n; + + if (s_copy) { + GGML_ASSERT(ggml_backend_buffer_is_host(s_copy->buffer)); + int32_t * data = (int32_t *) s_copy->data; + + // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n + for (uint32_t i = 0; i < n_kv; ++i) { + const uint32_t cell_id = i + kv_self->head; + + ////////////////////////////////////////////// + // TODO: this should not mutate the KV cache ! + llama_kv_cell & kv_cell = const_cast(kv_self)->cells[i]; + + // prevent out-of-bound sources + if (kv_cell.src < 0 || (uint32_t) kv_cell.src >= kv_self->size) { + kv_cell.src = cell_id; + } + + data[i] = kv_cell.src; + + // TODO: do not mutate the KV cache + // ensure copy only happens once + if (kv_cell.src != (int32_t) cell_id) { + kv_cell.src = cell_id; + } + } + } +} + +void llm_graph_input_s_mask::set_input(const llama_ubatch * ubatch) { + GGML_UNUSED(ubatch); + + const int64_t n_kv = kv_self->n; + + if (s_mask) { + GGML_ASSERT(ggml_backend_buffer_is_host(s_mask->buffer)); + float * data = (float *) s_mask->data; + + // clear unused states + for (int i = 0; i < n_kv; ++i) { + const uint32_t cell_id = i + kv_self->head; + + ////////////////////////////////////////////// + // TODO: this should not mutate the KV cache ! + llama_kv_cell & kv_cell = const_cast(kv_self)->cells[i]; + + data[i] = (float) (kv_cell.src >= 0); + + // only clear once + if (kv_cell.src < 0) { + kv_cell.src = cell_id; + } + } + } +} + +void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) { + GGML_UNUSED(ubatch); + + if (cross_embd && !cross->v_embd.empty()) { + assert(cross_embd->type == GGML_TYPE_F32); + + ggml_backend_tensor_set(cross_embd, cross->v_embd.data(), 0, ggml_nbytes(cross_embd)); + } +} + +void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) { + if (kq_mask) { + if (cparams.causal_attn) { + const int64_t n_kv = ubatch->n_tokens; + const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_seq_tokens = ubatch->n_seq_tokens; + const int64_t n_seqs = ubatch->n_seqs; + + GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer)); + float * data = (float *) kq_mask->data; + + for (int h = 0; h < 1; ++h) { + for (int s1 = 0; s1 < n_seqs; ++s1) { + const llama_seq_id seq_id = ubatch->seq_id[s1][0]; + + for (int j = 0; j < n_seq_tokens; ++j) { + const int32_t tj = s1*n_seq_tokens + j; + + for (int s0 = 0; s0 < n_seqs; ++s0) { + for (int i = 0; i < n_seq_tokens; ++i) { + const int32_t ti = s0*n_seq_tokens + i; + float f = -INFINITY; + + for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) { + if (ubatch->seq_id[s0][s] == seq_id && ubatch->pos[ti] <= ubatch->pos[tj]) { + if (hparams.use_alibi) { + f = -std::abs(ubatch->pos[ti] - ubatch->pos[tj]); + } else { + f = 0.0f; + } + break; + } + } + + data[h*(n_kv*n_tokens) + tj*n_kv + ti] = f; + } + } + } + } + } + } else { + const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_seq_tokens = ubatch->n_seq_tokens; + const int64_t n_seqs = ubatch->n_seqs; + const int64_t n_stride = ubatch->n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer)); + + float * data = (float *) kq_mask->data; + + for (int h = 0; h < 1; ++h) { + for (int s1 = 0; s1 < n_seqs; ++s1) { + const llama_seq_id seq_id = ubatch->seq_id[s1][0]; + + for (int j = 0; j < n_seq_tokens; ++j) { + const int32_t tj = s1*n_seq_tokens + j; + + for (int s0 = 0; s0 < n_seqs; ++s0) { + for (int i = 0; i < n_seq_tokens; ++i) { + const int32_t ti = s0*n_seq_tokens + i; + float f = -INFINITY; + + for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) { + if (ubatch->seq_id[s0][s] == seq_id) { + if (hparams.use_alibi) { + f = -std::abs(ubatch->pos[ti] - ubatch->pos[tj]); + } else { + f = 0.0f; + } + break; + } + } + + data[h*(n_tokens*n_tokens) + tj*n_stride + ti] = f; + } + } + + for (int i = n_tokens; i < n_stride; ++i) { + data[h*(n_tokens*n_tokens) + tj*n_stride + i] = -INFINITY; + } + } + } + } + } + } +} + +void llm_graph_input_attn_kv_unified::set_input(const llama_ubatch * ubatch) { + if (self_kq_mask || self_kq_mask_swa) { + // NOTE: hparams.causal_attn indicates the model is capable of generation and uses the kv cache. + if (cparams.causal_attn) { + const int64_t n_kv = kv_self->n; + const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_seq_tokens = ubatch->n_seq_tokens; + const int64_t n_seqs = ubatch->n_seqs; + + float * data = nullptr; + float * data_swa = nullptr; + + if (self_kq_mask) { + GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask->buffer)); + data = (float *) self_kq_mask->data; + } + + if (self_kq_mask_swa) { + GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask_swa->buffer)); + data_swa = (float *) self_kq_mask_swa->data; + } + + // For causal attention, use only the previous KV cells + // of the correct sequence for each token of the ubatch. + // It's assumed that if a token in the batch has multiple sequences, they are equivalent. + for (int h = 0; h < 1; ++h) { + for (int s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[s][0]; + + for (int j = 0; j < n_seq_tokens; ++j) { + const llama_pos pos = ubatch->pos[s*n_seq_tokens + j]; + + for (int i = 0; i < n_kv; ++i) { + float f; + if (!kv_self->cells[i].has_seq_id(seq_id) || kv_self->cells[i].pos > pos) { + f = -INFINITY; + } else { + if (hparams.use_alibi) { + f = -std::abs(kv_self->cells[i].pos - pos); + } else { + f = 0.0f; + } + } + + if (data) { + data[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f; + } + + // may need to cut off old tokens for sliding window + if (data_swa) { + if (pos - kv_self->cells[i].pos >= (int32_t)hparams.n_swa) { + f = -INFINITY; + } + data_swa[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f; + } + } + } + } + + if (data) { + for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { + for (int j = 0; j < n_kv; ++j) { + data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY; + } + } + } + + if (data_swa) { + for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { + for (int j = 0; j < n_kv; ++j) { + data_swa[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY; + } + } + } + } + } else { + const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_seq_tokens = ubatch->n_seq_tokens; + const int64_t n_seqs = ubatch->n_seqs; + // when using kv cache, the mask needs to match the kv cache size + const int64_t n_stride = n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask->buffer)); + + float * data = (float *) self_kq_mask->data; + + for (int h = 0; h < 1; ++h) { + for (int s1 = 0; s1 < n_seqs; ++s1) { + const llama_seq_id seq_id = ubatch->seq_id[s1][0]; + + for (int j = 0; j < n_seq_tokens; ++j) { + const int32_t tj = s1*n_seq_tokens + j; + + for (int s0 = 0; s0 < n_seqs; ++s0) { + for (int i = 0; i < n_seq_tokens; ++i) { + const int32_t ti = s0*n_seq_tokens + i; + float f = -INFINITY; + + for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) { + if (ubatch->seq_id[s0][s] == seq_id) { + if (hparams.use_alibi) { + f = -std::abs(ubatch->pos[ti] - ubatch->pos[tj]); + } else { + f = 0.0f; + } + break; + } + } + + data[h*(n_tokens*n_tokens) + tj*n_stride + ti] = f; + } + } + + for (int i = n_tokens; i < n_stride; ++i) { + data[h*(n_tokens*n_tokens) + tj*n_stride + i] = -INFINITY; + } + } + } + } + } + } +} + +void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) { + if (cross_kq_mask) { + const int64_t n_enc = cross_kq_mask->ne[0]; + const int64_t n_tokens = ubatch->n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(cross_kq_mask->buffer)); + GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing + + float * data = (float *) cross_kq_mask->data; + + for (int h = 0; h < 1; ++h) { + for (int j = 0; j < n_tokens; ++j) { + for (int i = 0; i < n_enc; ++i) { + float f = -INFINITY; + for (int s = 0; s < ubatch->n_seq_id[j]; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[j][s]; + if (cross->seq_ids_enc[i].find(seq_id) != cross->seq_ids_enc[i].end()) { + f = 0.0f; + } + } + data[h*(n_enc*n_tokens) + j*n_enc + i] = f; + } + } + + for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { + for (int j = 0; j < n_enc; ++j) { + data[h*(n_enc*n_tokens) + i*n_enc + j] = -INFINITY; + } + } + } + } +} + +// +// llm_graph_context +// + +llm_graph_context::llm_graph_context(const llm_graph_params & params) : + arch (params.arch), + hparams (params.hparams), + cparams (params.cparams), + ubatch (params.ubatch), + n_embd (hparams.n_embd), + n_layer (hparams.n_layer), + n_rot (hparams.n_rot), + n_ctx (cparams.n_ctx), + n_ctx_per_seq (cparams.n_ctx / cparams.n_seq_max), + n_head (hparams.n_head()), + n_head_kv (hparams.n_head_kv()), + n_embd_head_k (hparams.n_embd_head_k), + n_embd_k_gqa (hparams.n_embd_k_gqa()), + n_embd_head_v (hparams.n_embd_head_v), + n_embd_v_gqa (hparams.n_embd_v_gqa()), + n_expert (hparams.n_expert), + n_expert_used (hparams.n_expert_used), + freq_base (cparams.rope_freq_base), + freq_scale (cparams.rope_freq_scale), + ext_factor (cparams.yarn_ext_factor), + attn_factor (cparams.yarn_attn_factor), + beta_fast (cparams.yarn_beta_fast), + beta_slow (cparams.yarn_beta_slow), + norm_eps (hparams.f_norm_eps), + norm_rms_eps (hparams.f_norm_rms_eps), + n_tokens (ubatch.n_tokens), + n_outputs (params.n_outputs), + n_ctx_orig (cparams.n_ctx_orig_yarn), + pooling_type (cparams.pooling_type), + rope_type (hparams.rope_type), + ctx0 (params.ctx), + sched (params.sched), + backend_cpu (params.backend_cpu), + cvec (params.cvec), + loras (params.loras), + memory (params.memory), + cross (params.cross), + cb_func (params.cb), + res (std::make_unique()) { + } + +int64_t llm_graph_context::n_pos_per_token() const { + return arch == LLM_ARCH_QWEN2VL ? 4 : 1; +} + +void llm_graph_context::cb(ggml_tensor * cur, const char * name, int il) const { + if (cb_func) { + cb_func(ubatch, cur, name, il); + } +} + +ggml_tensor * llm_graph_context::build_cvec( + ggml_tensor * cur, + int il) const { + return cvec->apply_to(ctx0, cur, il); +} + +ggml_tensor * llm_graph_context::build_lora_mm( + ggml_tensor * w, + ggml_tensor * cur) const { + ggml_tensor * res = ggml_mul_mat(ctx0, w, cur); + + for (const auto & lora : *loras) { + llama_adapter_lora_weight * lw = lora.first->get_weight(w); + if (lw == nullptr) { + continue; + } + + const float adapter_scale = lora.second; + const float scale = lw->get_scale(lora.first->alpha, adapter_scale); + + ggml_tensor * ab_cur = ggml_mul_mat( + ctx0, lw->b, + ggml_mul_mat(ctx0, lw->a, cur) + ); + + ab_cur = ggml_scale(ctx0, ab_cur, scale); + res = ggml_add(ctx0, res, ab_cur); + } + + return res; +} + +ggml_tensor * llm_graph_context::build_lora_mm_id( + ggml_tensor * w, // ggml_tensor * as + ggml_tensor * cur, // ggml_tensor * b + ggml_tensor * ids) const { + ggml_tensor * res = ggml_mul_mat_id(ctx0, w, cur, ids); + for (const auto & lora : *loras) { + llama_adapter_lora_weight * lw = lora.first->get_weight(w); + if (lw == nullptr) { + continue; + } + + const float alpha = lora.first->alpha; + const float rank = (float) lw->b->ne[0]; + const float scale = alpha ? lora.second * alpha / rank : lora.second; + + ggml_tensor * ab_cur = ggml_mul_mat_id( + ctx0, lw->b, + ggml_mul_mat_id(ctx0, lw->a, cur, ids), + ids + ); + + ab_cur = ggml_scale(ctx0, ab_cur, scale); + res = ggml_add(ctx0, res, ab_cur); + } + + return res; +} + +ggml_tensor * llm_graph_context::build_norm( + ggml_tensor * cur, + ggml_tensor * mw, + ggml_tensor * mb, + llm_norm_type type, + int il) const { + switch (type) { + case LLM_NORM: cur = ggml_norm (ctx0, cur, hparams.f_norm_eps); break; + case LLM_NORM_RMS: cur = ggml_rms_norm(ctx0, cur, hparams.f_norm_rms_eps); break; + case LLM_NORM_GROUP: + { + cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], 1, cur->ne[1]); + cur = ggml_group_norm(ctx0, cur, hparams.n_norm_groups, hparams.f_norm_group_eps); + cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], cur->ne[2]); + } break; + } + + if (mw || mb) { + cb(cur, "norm", il); + } + + if (mw) { + cur = ggml_mul(ctx0, cur, mw); + if (mb) { + cb(cur, "norm_w", il); + } + } + + if (mb) { + cur = ggml_add(ctx0, cur, mb); + } + + return cur; +} + +ggml_tensor * llm_graph_context::build_ffn( + ggml_tensor * cur, + ggml_tensor * up, + ggml_tensor * up_b, + ggml_tensor * up_s, + ggml_tensor * gate, + ggml_tensor * gate_b, + ggml_tensor * gate_s, + ggml_tensor * down, + ggml_tensor * down_b, + ggml_tensor * down_s, + ggml_tensor * act_scales, + llm_ffn_op_type type_op, + llm_ffn_gate_type type_gate, + int il) const { + ggml_tensor * tmp = up ? build_lora_mm(up, cur) : cur; + cb(tmp, "ffn_up", il); + + if (up_b) { + tmp = ggml_add(ctx0, tmp, up_b); + cb(tmp, "ffn_up_b", il); + } + + if (up_s) { + tmp = ggml_mul(ctx0, tmp, up_s); + cb(tmp, "ffn_up_s", il); + } + + if (gate) { + switch (type_gate) { + case LLM_FFN_SEQ: + { + cur = build_lora_mm(gate, tmp); + cb(cur, "ffn_gate", il); + } break; + case LLM_FFN_PAR: + { + cur = build_lora_mm(gate, cur); + cb(cur, "ffn_gate", il); + } break; + } + + if (gate_b) { + cur = ggml_add(ctx0, cur, gate_b); + cb(cur, "ffn_gate_b", il); + } + + if (gate_s) { + cur = ggml_mul(ctx0, cur, gate_s); + cb(cur, "ffn_gate_s", il); + } + + } else { + cur = tmp; + } + + switch (type_op) { + case LLM_FFN_SILU: + { + cur = ggml_silu(ctx0, cur); + cb(cur, "ffn_silu", il); + } break; + case LLM_FFN_GELU: + { + cur = ggml_gelu(ctx0, cur); + cb(cur, "ffn_gelu", il); + if (act_scales != NULL) { + cur = ggml_div(ctx0, cur, act_scales); + cb(cur, "ffn_act", il); + } + } break; + case LLM_FFN_RELU: + { + cur = ggml_relu(ctx0, cur); + cb(cur, "ffn_relu", il); + } break; + case LLM_FFN_RELU_SQR: + { + cur = ggml_relu(ctx0, cur); + cb(cur, "ffn_relu", il); + + cur = ggml_sqr(ctx0, cur); + cb(cur, "ffn_sqr(relu)", il); + } break; + case LLM_FFN_SWIGLU: + { + // Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf + int64_t split_point = cur->ne[0] / 2; + ggml_tensor * x0 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], 0)); + ggml_tensor * x1 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur))); + + x0 = ggml_silu(ctx0, x0); + cb(cur, "ffn_silu", il); + + cur = ggml_mul(ctx0, x0, x1); + cb(cur, "ffn_mul", il); + } break; + } + + if (type_gate == LLM_FFN_PAR) { + cur = ggml_mul(ctx0, cur, tmp); + cb(cur, "ffn_gate_par", il); + } + + if (down) { + cur = build_lora_mm(down, cur); + } + + if (down_b) { + cb(cur, "ffn_down", il); + } + + if (down_b) { + cur = ggml_add(ctx0, cur, down_b); + } + + if (down_s) { + cur = ggml_mul(ctx0, cur, down_s); + cb(cur, "ffn_down_s", il); + } + + return cur; +} + +ggml_tensor * llm_graph_context::build_moe_ffn( + ggml_tensor * cur, + ggml_tensor * gate_inp, + ggml_tensor * up_exps, + ggml_tensor * gate_exps, + ggml_tensor * down_exps, + ggml_tensor * exp_probs_b, + int64_t n_expert, + int64_t n_expert_used, + llm_ffn_op_type type_op, + bool norm_w, + bool scale_w, + float w_scale, + llama_expert_gating_func_type gating_op, + int il) const { + int64_t n_embd = cur->ne[0]; + int64_t n_tokens = cur->ne[1]; + + ggml_tensor * logits = build_lora_mm(gate_inp, cur); // [n_expert, n_tokens] + cb(logits, "ffn_moe_logits", il); + + ggml_tensor * probs = nullptr; + switch (gating_op) { + case LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX: + { + probs = ggml_soft_max(ctx0, logits); // [n_expert, n_tokens] + } break; + case LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID: + { + probs = ggml_sigmoid(ctx0, logits); // [n_expert, n_tokens] + } break; + default: + GGML_ABORT("fatal error"); + } + cb(probs, "ffn_moe_probs", il); + + // add experts selection bias - introduced in DeepSeek V3 + // leave probs unbiased as it's later used to get expert weights + ggml_tensor * selection_probs = probs; + if (exp_probs_b != nullptr) { + selection_probs = ggml_add(ctx0, probs, exp_probs_b); + cb(selection_probs, "ffn_moe_probs_biased", il); + } + + // select experts + ggml_tensor * selected_experts = ggml_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens] + cb(selected_experts->src[0], "ffn_moe_argsort", il); + cb(selected_experts, "ffn_moe_topk", il); + + ggml_tensor * weights = ggml_get_rows(ctx0, + ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts); // [1, n_expert_used, n_tokens] + cb(weights, "ffn_moe_weights", il); + + if (norm_w) { + weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens); + + ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); // [1, n_tokens] + cb(weights_sum, "ffn_moe_weights_sum", il); + + weights = ggml_div(ctx0, weights, weights_sum); // [n_expert_used, n_tokens] + cb(weights, "ffn_moe_weights_norm", il); + + weights = ggml_reshape_3d(ctx0, weights, 1, n_expert_used, n_tokens); + } + if (scale_w) { + weights = ggml_scale(ctx0, weights, w_scale); + cb(weights, "ffn_moe_weights_scaled", il); + } + + cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens); + ggml_tensor * up = build_lora_mm_id(up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens] + cb(up, "ffn_moe_up", il); + + ggml_tensor * gate = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens] + cb(gate, "ffn_moe_gate", il); + + switch (type_op) { + case LLM_FFN_SILU: + { + gate = ggml_silu(ctx0, gate); + cb(gate, "ffn_moe_silu", il); + } break; + case LLM_FFN_GELU: + { + gate = ggml_gelu(ctx0, gate); + cb(gate, "ffn_moe_gelu", il); + } break; + default: + GGML_ABORT("fatal error"); + } + + ggml_tensor * par = ggml_mul(ctx0, up, gate); // [n_ff, n_expert_used, n_tokens] + cb(par, "ffn_moe_gate_par", il); + + ggml_tensor * experts = build_lora_mm_id(down_exps, par, selected_experts); // [n_embd, n_expert_used, n_tokens] + cb(experts, "ffn_moe_down", il); + + experts = ggml_mul(ctx0, experts, weights); + + // aggregate experts + ggml_tensor * moe_out = nullptr; + for (int i = 0; i < n_expert_used; ++i) { + ggml_tensor * cur_expert = ggml_view_2d(ctx0, experts, n_embd, n_tokens, + experts->nb[2], i*experts->nb[1]); + + if (i == 0) { + moe_out = cur_expert; + } else { + moe_out = ggml_add(ctx0, moe_out, cur_expert); + } + } + + if (n_expert_used == 1) { + // avoid returning a non-contiguous tensor + moe_out = ggml_cont(ctx0, moe_out); + } + + return moe_out; +} + +// input embeddings with optional lora +ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const { + const int64_t n_embd = hparams.n_embd; + + auto inp = std::make_unique(); + + ggml_tensor * cur = nullptr; + + if (ubatch.token) { + inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens); + //cb(inp->tokens, "inp_tokens", -1); + ggml_set_input(inp->tokens); + + cur = ggml_get_rows(ctx0, tok_embd, inp->tokens); + + // apply lora for embedding tokens if needed + for (const auto & lora : *loras) { + llama_adapter_lora_weight * lw = lora.first->get_weight(tok_embd); + if (lw == nullptr) { + continue; + } + + const float adapter_scale = lora.second; + const float scale = lw->get_scale(lora.first->alpha, adapter_scale); + + ggml_tensor * inpL_delta = ggml_scale(ctx0, ggml_mul_mat( + ctx0, lw->b, // non-transposed lora_b + ggml_get_rows(ctx0, lw->a, inp->tokens) + ), scale); + + cur = ggml_add(ctx0, cur, inpL_delta); + } + } else { + inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, ubatch.n_tokens); + ggml_set_input(inp->embd); + + cur = inp->embd; + } + + // For Granite architecture + if (hparams.f_embedding_scale != 0.0f) { + cur = ggml_scale(ctx0, cur, hparams.f_embedding_scale); + } + + cb(cur, "inp_embd", -1); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_pos() const { + auto inp = std::make_unique(n_pos_per_token()); + + auto & cur = inp->pos; + + cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens*n_pos_per_token()); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_out_ids() const { + auto inp = std::make_unique(hparams, cparams, n_outputs); + + auto & cur = inp->out_ids; + + cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_outputs); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_mean() const { + auto inp = std::make_unique(cparams); + + auto & cur = inp->mean; + + cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_cls() const { + auto inp = std::make_unique(cparams); + + auto & cur = inp->cls; + + cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_s_copy() const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + auto inp = std::make_unique(kv_self); + + const auto n_kv = kv_self->n; + + auto & cur = inp->s_copy; + + cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_kv); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_s_mask() const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + auto inp = std::make_unique(kv_self); + + const auto n_kv = kv_self->n; + + auto & cur = inp->s_mask; + + cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, n_kv); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_cross_embd() const { + auto inp = std::make_unique(cross); + + auto & cur = inp->cross_embd; + + // if we have the output embeddings from the encoder, use them directly + // TODO: needs more work to be correct, for now just use the tensor shape + //if (cross->t_embd) { + // cur = ggml_view_tensor(ctx0, cross->t_embd); + + // return cur; + //} + + const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd; + const auto n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train; + + cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_enc); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_pos_bucket_enc() const { + auto inp = std::make_unique(hparams); + + auto & cur = inp->pos_bucket; + + cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_tokens, n_tokens); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_pos_bucket_dec() const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + auto inp = std::make_unique(hparams, kv_self); + + const auto n_kv = kv_self->n; + + auto & cur = inp->pos_bucket; + + cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_kv, n_tokens); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_pos_bias(ggml_tensor * pos_bucket, ggml_tensor * attn_rel_b) const { + ggml_tensor * pos_bucket_1d = ggml_reshape_1d(ctx0, pos_bucket, pos_bucket->ne[0] * pos_bucket->ne[1]); + cb(pos_bucket_1d, "pos_bucket_1d", -1); + + ggml_tensor * pos_bias = ggml_get_rows(ctx0, attn_rel_b, pos_bucket_1d); + + pos_bias = ggml_reshape_3d(ctx0, pos_bias, pos_bias->ne[0], pos_bucket->ne[0], pos_bucket->ne[1]); + pos_bias = ggml_permute (ctx0, pos_bias, 2, 0, 1, 3); + pos_bias = ggml_cont (ctx0, pos_bias); + + cb(pos_bias, "pos_bias", -1); + + return pos_bias; +} + +ggml_tensor * llm_graph_context::build_attn_mha( + ggml_cgraph * gf, + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * kq_b, + ggml_tensor * kq_mask, + bool v_trans, + float kq_scale) const { + //const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); + //const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); + + //const int64_t n_head = hparams.n_head(il); + //const int64_t n_head_kv = hparams.n_head_kv(il); + + //const auto & n_embd_head_k = hparams.n_embd_head_k; + //const auto & n_embd_head_v = hparams.n_embd_head_v; + + const auto n_embd_head_v = v_trans ? v->ne[1] : v->ne[0]; + + const auto n_tokens = q->ne[1]; + const auto n_head = q->ne[2]; + const auto n_kv = k->ne[1]; + + ggml_tensor * cur; + + // TODO: replace hardcoded padding with ggml-provided padding + if (cparams.flash_attn && (n_kv % 256 == 0) && kq_b == nullptr) { + GGML_ASSERT(kq_b == nullptr && "Flash attention does not support KQ bias yet"); + + if (v_trans) { + v = ggml_transpose(ctx0, v); + } + + cur = ggml_flash_attn_ext(ctx0, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias, + hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f); + + ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32); + + cur = ggml_reshape_2d(ctx0, cur, n_embd_head_v*n_head, n_tokens); + } else { + ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + + // note: this op tends to require high floating point range + // while for some models F16 is enough, for others it is not, so we default to F32 here + ggml_mul_mat_set_prec(kq, GGML_PREC_F32); + + if (arch == LLM_ARCH_GROK) { + // need to do the following: + // multiply by attn_output_multiplyer of 0.08838834764831845 + // and then : + // kq = 30 * tanh(kq / 30) + // before the softmax below + + kq = ggml_tanh(ctx0, ggml_scale(ctx0, kq, 0.08838834764831845f/30.0f)); + kq = ggml_scale(ctx0, kq, 30); + } + + if (hparams.attn_soft_cap) { + kq = ggml_scale(ctx0, kq, 1.0f / hparams.f_attn_logit_softcapping); + kq = ggml_tanh (ctx0, kq); + kq = ggml_scale(ctx0, kq, hparams.f_attn_logit_softcapping); + } + + if (kq_b) { + kq = ggml_add(ctx0, kq, kq_b); + } + + kq = ggml_soft_max_ext(ctx0, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias); + + if (!v_trans) { + // note: avoid this branch + v = ggml_cont(ctx0, ggml_transpose(ctx0, v)); + } + + ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq); + + ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); + + cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_head_v*n_head, n_tokens); + + if (!cparams.offload_kqv) { + // all nodes between the KV store and the attention output are run on the CPU + ggml_backend_sched_set_tensor_backend(sched, cur, backend_cpu); + } + } + + ggml_build_forward_expand(gf, cur); + + return cur; +} + +llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() const { + auto inp = std::make_unique(hparams, cparams); + + // note: there is no KV cache, so the number of KV values is equal to the number of tokens in the batch + inp->kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); + //cb(inp_kq_mask, "KQ_mask", -1); + ggml_set_input(inp->kq_mask); + + inp->kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->kq_mask, GGML_TYPE_F16) : inp->kq_mask; + + return (llm_graph_input_attn_no_cache *) res->add_input(std::move(inp)); +} + +ggml_tensor * llm_graph_context::build_attn( + llm_graph_input_attn_no_cache * inp, + ggml_cgraph * gf, + ggml_tensor * wo, + ggml_tensor * wo_b, + ggml_tensor * q_cur, + ggml_tensor * k_cur, + ggml_tensor * v_cur, + ggml_tensor * kq_b, + float kq_scale, + int il) const { + GGML_UNUSED(n_tokens); + + // these nodes are added to the graph together so that they are not reordered + // by doing so, the number of splits in the graph is reduced + ggml_build_forward_expand(gf, q_cur); + ggml_build_forward_expand(gf, k_cur); + ggml_build_forward_expand(gf, v_cur); + + const auto & kq_mask = inp->get_kq_mask(); + + ggml_tensor * q = ggml_permute(ctx0, q_cur, 0, 2, 1, 3); + //cb(q, "q", il); + + ggml_tensor * k = ggml_permute(ctx0, k_cur, 0, 2, 1, 3); + //cb(k, "k", il); + + ggml_tensor * v = ggml_permute(ctx0, v_cur, 0, 2, 1, 3); + //cb(k, "v", il); + + ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, false, kq_scale); + + cb(cur, "kqv_out", il); + + if (wo) { + cur = build_lora_mm(wo, cur); + } + + if (wo_b) { + //cb(cur, "kqv_wo", il); + } + + if (wo_b) { + cur = ggml_add(ctx0, cur, wo_b); + } + + return cur; +} + +llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified( + bool causal, + bool swa) const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + auto inp = std::make_unique(hparams, cparams, kv_self); + + const auto n_kv = kv_self->n; + + inp->self_kq_mask = causal + ? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)) + : ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); + //cb(inp->self_kq_mask, "KQ_mask", -1); + ggml_set_input(inp->self_kq_mask); + + inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask; + + if (swa) { + GGML_ASSERT(hparams.n_swa > 0); + + inp->self_kq_mask_swa = causal + ? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)) + : ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); + //cb(inp->self_kq_mask_swa, "KQ_mask_swa", -1); + ggml_set_input(inp->self_kq_mask_swa); + + inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa; + } + + return (llm_graph_input_attn_kv_unified *) res->add_input(std::move(inp)); +} + +ggml_tensor * llm_graph_context::build_attn( + llm_graph_input_attn_kv_unified * inp, + ggml_cgraph * gf, + ggml_tensor * wo, + ggml_tensor * wo_b, + ggml_tensor * q_cur, + ggml_tensor * k_cur, + ggml_tensor * v_cur, + ggml_tensor * kq_b, + float kq_scale, + int il) const { + // these nodes are added to the graph together so that they are not reordered + // by doing so, the number of splits in the graph is reduced + ggml_build_forward_expand(gf, q_cur); + ggml_build_forward_expand(gf, k_cur); + ggml_build_forward_expand(gf, v_cur); + + const llama_kv_cache_unified * kv_self = static_cast(memory); + const auto & n_ctx = cparams.n_ctx; + + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); + const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); + + const auto n_tokens = q_cur->ne[2]; + + const bool v_trans = !cparams.flash_attn; + + // store to KV cache + { + GGML_ASSERT(!kv_self->recurrent); + + const auto kv_head = kv_self->head; + + GGML_ASSERT(kv_self->size == n_ctx); + + ggml_tensor * k_cache_view = ggml_view_1d(ctx0, kv_self->k_l[il], n_tokens*n_embd_k_gqa, ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa)*kv_head); + //cb(k_cache_view, "k_cache_view", il); + + // note: storing RoPE-ed version of K in the KV cache + ggml_build_forward_expand(gf, ggml_cpy(ctx0, k_cur, k_cache_view)); + + assert(v_cur->ne[0] == n_embd_v_gqa && v_cur->ne[1] == n_tokens); + + ggml_tensor * v_cache_view = nullptr; + + if (!v_trans) { + v_cache_view = ggml_view_1d(ctx0, kv_self->v_l[il], n_tokens*n_embd_v_gqa, ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa)*kv_head); + } else { + // note: the V cache is transposed when not using flash attention + v_cache_view = ggml_view_2d(ctx0, kv_self->v_l[il], n_tokens, n_embd_v_gqa, + ( n_ctx)*ggml_element_size(kv_self->v_l[il]), + (kv_head)*ggml_element_size(kv_self->v_l[il])); + + v_cur = ggml_transpose(ctx0, v_cur); + } + //cb(v_cache_view, "v_cache_view", il); + + ggml_build_forward_expand(gf, ggml_cpy(ctx0, v_cur, v_cache_view)); + } + + const bool is_sliding = hparams.is_sliding(il); + + const auto & kq_mask = is_sliding ? inp->get_kq_mask_swa() : inp->get_kq_mask(); + + const auto n_kv = kv_self->n; + + const int64_t n_head_kv = hparams.n_head_kv(il); + + const auto & n_embd_head_k = hparams.n_embd_head_k; + const auto & n_embd_head_v = hparams.n_embd_head_v; + + ggml_tensor * q = ggml_permute(ctx0, q_cur, 0, 2, 1, 3); + //cb(q, "q", il); + + ggml_tensor * k = + ggml_view_3d(ctx0, kv_self->k_l[il], + n_embd_head_k, n_kv, n_head_kv, + ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa), + ggml_row_size(kv_self->k_l[il]->type, n_embd_head_k), + 0); + //cb(k, "k", il); + + ggml_tensor * v = !v_trans ? + ggml_view_3d(ctx0, kv_self->v_l[il], + n_embd_head_v, n_kv, n_head_kv, + ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa), + ggml_row_size(kv_self->v_l[il]->type, n_embd_head_v), + 0) : + ggml_view_3d(ctx0, kv_self->v_l[il], + n_kv, n_embd_head_v, n_head_kv, + ggml_element_size(kv_self->v_l[il])*n_ctx, + ggml_element_size(kv_self->v_l[il])*n_ctx*n_embd_head_v, + 0); + + ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_trans, kq_scale); + cb(cur, "kqv_out", il); + + if (wo) { + cur = build_lora_mm(wo, cur); + } + + if (wo_b) { + //cb(cur, "kqv_wo", il); + } + + if (wo_b) { + cur = ggml_add(ctx0, cur, wo_b); + } + + return cur; +} + +llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const { + auto inp = std::make_unique(cross); + + const int32_t n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train; + + inp->cross_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); + ggml_set_input(inp->cross_kq_mask); + + inp->cross_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->cross_kq_mask, GGML_TYPE_F16) : inp->cross_kq_mask; + + return (llm_graph_input_attn_cross *) res->add_input(std::move(inp)); +} + +ggml_tensor * llm_graph_context::build_attn( + llm_graph_input_attn_cross * inp, + ggml_cgraph * gf, + ggml_tensor * wo, + ggml_tensor * wo_b, + ggml_tensor * q_cur, + ggml_tensor * k_cur, + ggml_tensor * v_cur, + ggml_tensor * kq_b, + float kq_scale, + int il) const { + // these nodes are added to the graph together so that they are not reordered + // by doing so, the number of splits in the graph is reduced + ggml_build_forward_expand(gf, q_cur); + ggml_build_forward_expand(gf, k_cur); + ggml_build_forward_expand(gf, v_cur); + + const auto & kq_mask = inp->get_kq_mask_cross(); + + ggml_tensor * q = ggml_permute(ctx0, q_cur, 0, 2, 1, 3); + //cb(q, "q", il); + + ggml_tensor * k = ggml_permute(ctx0, k_cur, 0, 2, 1, 3); + //cb(k, "k", il); + + ggml_tensor * v = ggml_permute(ctx0, v_cur, 0, 2, 1, 3); + //cb(k, "v", il); + + ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, false, kq_scale); + + cb(cur, "kqv_out", il); + + if (wo) { + cur = build_lora_mm(wo, cur); + } + + if (wo_b) { + //cb(cur, "kqv_wo", il); + } + + if (wo_b) { + cur = ggml_add(ctx0, cur, wo_b); + } + + return cur; +} + +ggml_tensor * llm_graph_context::build_copy_mask_state( + ggml_cgraph * gf, + ggml_tensor * s, + ggml_tensor * state_copy, + ggml_tensor * state_mask, + int32_t n_state, + int32_t n_seqs) const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + const auto n_kv = kv_self->n; + const auto kv_head = kv_self->head; + + ggml_tensor * states = ggml_reshape_2d(ctx0, s, n_state, kv_self->size); + + // copy states + // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv + // this shrinks the tensors's ne[1] to n_kv + states = ggml_get_rows(ctx0, states, state_copy); + + // clear states of sequences which are starting at the beginning of this batch + // FIXME: zero-out NANs? + states = ggml_mul(ctx0, states, state_mask); + + // copy states which won't be changed further (between n_seqs and n_kv) + ggml_build_forward_expand(gf, + ggml_cpy(ctx0, + ggml_view_1d(ctx0, states, n_state*(n_kv - n_seqs), (n_seqs )*n_state*ggml_element_size(states)), + ggml_view_1d(ctx0, s, n_state*(n_kv - n_seqs), (kv_head + n_seqs)*n_state*ggml_element_size(s)))); + + // the part of the states that will be used and modified + return ggml_view_2d(ctx0, states, n_state, n_seqs, states->nb[1], 0); +} + +ggml_tensor * llm_graph_context::build_rwkv_token_shift_load( + ggml_cgraph * gf, + ggml_tensor * state_copy, + ggml_tensor * state_mask, + const llama_ubatch & ubatch, + int il) const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + const auto token_shift_count = hparams.token_shift_count; + + const int64_t n_seqs = ubatch.n_seqs; + + ggml_tensor * token_shift_all = kv_self->k_l[il]; + + ggml_tensor * token_shift = build_copy_mask_state( + gf, token_shift_all, state_copy, state_mask, + hparams.n_embd_k_s(), n_seqs); + + token_shift = ggml_reshape_3d(ctx0, token_shift, hparams.n_embd, token_shift_count, n_seqs); + + return token_shift; +} + +ggml_tensor * llm_graph_context::build_rwkv_token_shift_store( + ggml_tensor * token_shift, + const llama_ubatch & ubatch, + int il) const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + const auto token_shift_count = hparams.token_shift_count; + const auto n_embd = hparams.n_embd; + + const int64_t n_seqs = ubatch.n_seqs; + + const auto kv_head = kv_self->head; + + return ggml_cpy( + ctx0, + ggml_view_1d(ctx0, token_shift, n_embd * n_seqs * token_shift_count, 0), + ggml_view_1d(ctx0, kv_self->k_l[il], hparams.n_embd_k_s() * n_seqs, hparams.n_embd_k_s() * kv_head * ggml_element_size(kv_self->k_l[il])) + ); +} + +void llm_graph_context::build_pooling( + ggml_cgraph * gf, + ggml_tensor * cls, + ggml_tensor * cls_b, + ggml_tensor * cls_out, + ggml_tensor * cls_out_b) const { + if (!cparams.embeddings) { + return; + } + + ggml_tensor * inp = res->t_embd; + + //// find result_norm tensor for input + //for (int i = ggml_graph_n_nodes(gf) - 1; i >= 0; --i) { + // inp = ggml_graph_node(gf, i); + // if (strcmp(inp->name, "result_norm") == 0 || strcmp(inp->name, "result_embd") == 0) { + // break; + // } + + // inp = nullptr; + //} + + GGML_ASSERT(inp != nullptr && "missing result_norm/result_embd tensor"); + + ggml_tensor * cur; + + switch (pooling_type) { + case LLAMA_POOLING_TYPE_NONE: + { + cur = inp; + } break; + case LLAMA_POOLING_TYPE_MEAN: + { + ggml_tensor * inp_mean = build_inp_mean(); + cur = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, inp)), inp_mean); + } break; + case LLAMA_POOLING_TYPE_CLS: + case LLAMA_POOLING_TYPE_LAST: + { + ggml_tensor * inp_cls = build_inp_cls(); + cur = ggml_get_rows(ctx0, inp, inp_cls); + } break; + case LLAMA_POOLING_TYPE_RANK: + { + ggml_tensor * inp_cls = build_inp_cls(); + inp = ggml_get_rows(ctx0, inp, inp_cls); + + // classification head + // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566 + GGML_ASSERT(cls != nullptr); + GGML_ASSERT(cls_b != nullptr); + + cur = ggml_add (ctx0, ggml_mul_mat(ctx0, cls, inp), cls_b); + cur = ggml_tanh(ctx0, cur); + + // some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en + // https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896 + if (cls_out) { + GGML_ASSERT(cls_out_b != nullptr); + + cur = ggml_add (ctx0, ggml_mul_mat(ctx0, cls_out, cur), cls_out_b); + } + } break; + default: + { + GGML_ABORT("unknown pooling type"); + } + } + + cb(cur, "result_embd_pooled", -1); + res->t_embd_pooled = cur; + + ggml_build_forward_expand(gf, cur); +} + diff --git a/src/llama-graph.h b/src/llama-graph.h new file mode 100644 index 0000000000..b7a66d1898 --- /dev/null +++ b/src/llama-graph.h @@ -0,0 +1,576 @@ +#pragma once + +#include "llama-arch.h" +#include "llama-hparams.h" +#include "llama-adapter.h" + +#include +#include +#include +#include +#include + +struct ggml_cgraph; +struct ggml_context; +struct ggml_tensor; + +struct llama_ubatch; +struct llama_cparams; + +class llama_memory_i; +class llama_kv_cache_unified; + +// certain models (typically multi-modal) can produce different types of graphs +enum llm_graph_type { + LLM_GRAPH_TYPE_DEFAULT, + LLM_GRAPH_TYPE_ENCODER, + LLM_GRAPH_TYPE_DECODER, +}; + +enum llm_ffn_op_type { + LLM_FFN_SILU, + LLM_FFN_GELU, + LLM_FFN_RELU, + LLM_FFN_RELU_SQR, + LLM_FFN_SWIGLU, +}; + +enum llm_ffn_gate_type { + LLM_FFN_SEQ, + LLM_FFN_PAR, // ffn_gate is parallel to ffn_up +}; + +enum llm_norm_type { + LLM_NORM, + LLM_NORM_RMS, + LLM_NORM_GROUP, +}; + +// TODO: tmp - need something better to pass the data from the encoder to the decoder +struct llama_cross { + // the output embeddings from the encoder as a ggml tensor + // TODO: this needs more work to be correct, for now copy the embeddings data to host memory + // ref: https://github.com/ggml-org/llama.cpp/pull/11213#discussion_r1969892524 + //ggml_tensor * t_embd = nullptr; + + int64_t n_embd = 0; + int64_t n_enc = 0; + + // embeddings data copied to host memory (tmp) + std::vector v_embd; + + // needed to construct the cross-attention mask in the decoder + std::vector> seq_ids_enc; +}; + +// +// llm_graph_input +// + +class llm_graph_input_i { +public: + virtual ~llm_graph_input_i() = default; + + virtual void set_input(const llama_ubatch * ubatch) = 0; +}; + +using llm_graph_input_ptr = std::unique_ptr; + + +class llm_graph_input_embd : public llm_graph_input_i { +public: + llm_graph_input_embd() = default; + virtual ~llm_graph_input_embd() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * tokens = nullptr; // I32 [n_batch] + ggml_tensor * embd = nullptr; // F32 [n_embd, n_batch] +}; + +class llm_graph_input_pos : public llm_graph_input_i { +public: + llm_graph_input_pos(int64_t n_pos_per_token) : n_pos_per_token(n_pos_per_token) {} + virtual ~llm_graph_input_pos() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * pos = nullptr; // I32 [n_batch] + + const int64_t n_pos_per_token = 1; +}; + +class llm_graph_input_pos_bucket : public llm_graph_input_i { +public: + llm_graph_input_pos_bucket(const llama_hparams & hparams) : hparams(hparams) {} + virtual ~llm_graph_input_pos_bucket() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * pos_bucket = nullptr; // I32 [n_batch, n_batch] + + const llama_hparams & hparams; +}; + +class llm_graph_input_pos_bucket_kv : public llm_graph_input_i { +public: + llm_graph_input_pos_bucket_kv( + const llama_hparams & hparams, + const llama_kv_cache_unified * kv_self) : hparams(hparams), kv_self(kv_self) {} + virtual ~llm_graph_input_pos_bucket_kv() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * pos_bucket = nullptr; // I32 [n_kv, n_batch] + + const llama_hparams & hparams; + const llama_kv_cache_unified * kv_self; +}; + +class llm_graph_input_out_ids : public llm_graph_input_i { +public: + llm_graph_input_out_ids( + const llama_hparams & hparams, + const llama_cparams & cparams, + int32_t n_outputs) : hparams(hparams), cparams(cparams), n_outputs(n_outputs) {} + virtual ~llm_graph_input_out_ids() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * out_ids; // I32 [n_outputs] + + const llama_hparams & hparams; + const llama_cparams & cparams; + + const int32_t n_outputs; +}; + +class llm_graph_input_mean : public llm_graph_input_i { +public: + llm_graph_input_mean(const llama_cparams & cparams) : cparams(cparams) {} + virtual ~llm_graph_input_mean() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * mean; // F32 [n_batch, n_batch] + + const llama_cparams & cparams; +}; + +class llm_graph_input_cls : public llm_graph_input_i { +public: + llm_graph_input_cls(const llama_cparams & cparams) : cparams(cparams) {} + virtual ~llm_graph_input_cls() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * cls; // I32 [n_batch] + + const llama_cparams & cparams; +}; + +class llm_graph_input_s_copy : public llm_graph_input_i { +public: + llm_graph_input_s_copy(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {} + virtual ~llm_graph_input_s_copy() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * s_copy; // I32 [kv_size] + + const llama_kv_cache_unified * kv_self; +}; + +class llm_graph_input_s_mask : public llm_graph_input_i { +public: + llm_graph_input_s_mask(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {} + virtual ~llm_graph_input_s_mask() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * s_mask; // F32 [1, n_kv] + + const llama_kv_cache_unified * kv_self; +}; + +class llm_graph_input_cross_embd : public llm_graph_input_i { +public: + llm_graph_input_cross_embd( + const llama_cross * cross) : cross(cross) {} + virtual ~llm_graph_input_cross_embd() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * cross_embd; // F32 [n_embd, n_outputs_enc] + + const llama_cross * cross; +}; + +class llm_graph_input_attn_no_cache : public llm_graph_input_i { +public: + llm_graph_input_attn_no_cache(const llama_hparams & hparams, const llama_cparams & cparams) : + hparams(hparams), + cparams(cparams) { + } + ~llm_graph_input_attn_no_cache() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * get_kq_mask() const { return kq_mask_cnv; } + + ggml_tensor * kq_mask = nullptr; // F32 [n_tokens, n_batch] + ggml_tensor * kq_mask_cnv = nullptr; // [n_tokens, n_batch] + + const llama_hparams & hparams; + const llama_cparams & cparams; +}; + +class llm_graph_input_attn_kv_unified : public llm_graph_input_i { +public: + llm_graph_input_attn_kv_unified( + const llama_hparams & hparams, + const llama_cparams & cparams, + const llama_kv_cache_unified * kv_self) : + hparams(hparams), + cparams(cparams), + kv_self(kv_self) { + } + ~llm_graph_input_attn_kv_unified() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; } + ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; } + + ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch] + ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch] + ggml_tensor * self_kq_mask_swa = nullptr; // F32 [n_kv, n_batch] + ggml_tensor * self_kq_mask_swa_cnv = nullptr; // [n_kv, n_batch] + + const llama_hparams & hparams; + const llama_cparams & cparams; + + const llama_kv_cache_unified * kv_self; +}; + +class llm_graph_input_attn_cross : public llm_graph_input_i { +public: + llm_graph_input_attn_cross(const llama_cross * cross) : cross(cross) {} + ~llm_graph_input_attn_cross() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * get_kq_mask_cross() const { return cross_kq_mask_cnv; } + + ggml_tensor * cross_kq_mask = nullptr; // F32 [n_outputs_enc, n_batch] + ggml_tensor * cross_kq_mask_cnv = nullptr; // F32 [n_outputs_enc, n_batch] + + const llama_cross * cross = nullptr; +}; + +// +// llm_graph_result +// + +// these objects deliver the result from the graph build process back to the llama_context +// note that the input tensors created for the graph are referenced here - the goal is to be able to populate their +// specific data, by calling the set_inputs() method +// along with the input tensors, the object also provides commonly used outputs tensors, such as logits, embeddings, etc. +// these are used by the llama_context to extact the relevant data, based on the compute parameters + +class llm_graph_result_i { +public: + virtual ~llm_graph_result_i() = default; + + virtual ggml_tensor * get_logits() = 0; + virtual ggml_tensor * get_embd() = 0; + virtual ggml_tensor * get_embd_pooled() = 0; + + virtual void set_inputs(const llama_ubatch * ubatch) = 0; +}; + +using llm_graph_result_ptr = std::unique_ptr; + + +class llm_graph_result : public llm_graph_result_i { +public: + virtual ~llm_graph_result() = default; + + ggml_tensor * get_logits() override { return t_logits; } + ggml_tensor * get_embd() override { return t_embd; } + ggml_tensor * get_embd_pooled() override { return t_embd_pooled; } + + void set_inputs(const llama_ubatch * ubatch) override { + for (auto & input : inputs) { + input->set_input(ubatch); + } + } + + llm_graph_input_i * add_input(llm_graph_input_ptr input) { + inputs.emplace_back(std::move(input)); + return inputs.back().get(); + } + + // important graph nodes + ggml_tensor * t_logits = nullptr; + ggml_tensor * t_embd = nullptr; + ggml_tensor * t_embd_pooled = nullptr; + + std::vector inputs; +}; + +// +// llm_graph_context +// + +// callback that allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.) +using llm_graph_cb = std::function; + +struct llm_graph_params { + ggml_context * ctx; + + const llm_arch arch; + + const llama_hparams & hparams; + const llama_cparams & cparams; + const llama_ubatch & ubatch; + + ggml_backend_sched * sched; + ggml_backend * backend_cpu; + + const llama_adapter_cvec * cvec; + const llama_adapter_loras * loras; + const llama_memory_i * memory; + const llama_cross * cross; + + int32_t n_outputs; + + const llm_graph_cb & cb; +}; + +struct llm_graph_context { + const llm_arch arch; + + const llama_hparams & hparams; + const llama_cparams & cparams; + const llama_ubatch & ubatch; + + const int64_t n_embd; + const int64_t n_layer; + const int64_t n_rot; + const int64_t n_ctx; // user-specified context size (can be different from n_ctx_train) + const int64_t n_ctx_per_seq; + const int64_t n_head; + const int64_t n_head_kv; + const int64_t n_embd_head_k; + const int64_t n_embd_k_gqa; + const int64_t n_embd_head_v; + const int64_t n_embd_v_gqa; + const int64_t n_expert; + const int64_t n_expert_used; + + const float freq_base; + const float freq_scale; + const float ext_factor; + const float attn_factor; + const float beta_fast; + const float beta_slow; + const float norm_eps; + const float norm_rms_eps; + + const int32_t n_tokens; + const int32_t n_outputs; + const int32_t n_ctx_orig; // yarn + + const enum llama_pooling_type pooling_type; + const enum llama_rope_type rope_type; + + ggml_context * ctx0 = nullptr; + + ggml_backend_sched * sched; + + ggml_backend * backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove? + + const llama_adapter_cvec * cvec; + const llama_adapter_loras * loras; + const llama_memory_i * memory; + const llama_cross * cross; + + const llm_graph_cb & cb_func; + + std::unique_ptr res; + + llm_graph_context(const llm_graph_params & params); + + int64_t n_pos_per_token() const; + + void cb(ggml_tensor * cur, const char * name, int il) const; + + // + // common + // + + ggml_tensor * build_cvec( + ggml_tensor * cur, + int il) const; + + // do mat_mul, while optionally apply lora + ggml_tensor * build_lora_mm( + ggml_tensor * w, + ggml_tensor * cur) const; + + // do mat_mul_id, while optionally apply lora + ggml_tensor * build_lora_mm_id( + ggml_tensor * w, // ggml_tensor * as + ggml_tensor * cur, // ggml_tensor * b + ggml_tensor * ids) const; + + ggml_tensor * build_norm( + ggml_tensor * cur, + ggml_tensor * mw, + ggml_tensor * mb, + llm_norm_type type, + int il) const; + + ggml_tensor * build_ffn( + ggml_tensor * cur, + ggml_tensor * up, + ggml_tensor * up_b, + ggml_tensor * up_s, + ggml_tensor * gate, + ggml_tensor * gate_b, + ggml_tensor * gate_s, + ggml_tensor * down, + ggml_tensor * down_b, + ggml_tensor * down_s, + ggml_tensor * act_scales, + llm_ffn_op_type type_op, + llm_ffn_gate_type type_gate, + int il) const; + + ggml_tensor * build_moe_ffn( + ggml_tensor * cur, + ggml_tensor * gate_inp, + ggml_tensor * up_exps, + ggml_tensor * gate_exps, + ggml_tensor * down_exps, + ggml_tensor * exp_probs_b, + int64_t n_expert, + int64_t n_expert_used, + llm_ffn_op_type type_op, + bool norm_w, + bool scale_w, + float w_scale, + llama_expert_gating_func_type gating_op, + int il) const; + + // + // inputs + // + + ggml_tensor * build_inp_embd(ggml_tensor * tok_embd) const; + ggml_tensor * build_inp_pos() const; + ggml_tensor * build_inp_out_ids() const; + ggml_tensor * build_inp_mean() const; + ggml_tensor * build_inp_cls() const; + ggml_tensor * build_inp_s_copy() const; + ggml_tensor * build_inp_s_mask() const; + + ggml_tensor * build_inp_cross_embd() const; + ggml_tensor * build_inp_pos_bucket_enc() const; + ggml_tensor * build_inp_pos_bucket_dec() const; + ggml_tensor * build_pos_bias(ggml_tensor * pos_bucket, ggml_tensor * attn_rel_b) const; + + // + // attention + // + + ggml_tensor * build_attn_mha( + ggml_cgraph * gf, + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * kq_b, + ggml_tensor * kq_mask, + bool v_trans, + float kq_scale) const; + + llm_graph_input_attn_no_cache * build_attn_inp_no_cache() const; + + ggml_tensor * build_attn( + llm_graph_input_attn_no_cache * inp, + ggml_cgraph * gf, + ggml_tensor * wo, + ggml_tensor * wo_b, + ggml_tensor * q_cur, + ggml_tensor * k_cur, + ggml_tensor * v_cur, + ggml_tensor * kq_b, + float kq_scale, + int il) const; + + llm_graph_input_attn_kv_unified * build_attn_inp_kv_unified( + bool causal, + bool swa) const; + + ggml_tensor * build_attn( + llm_graph_input_attn_kv_unified * inp, + ggml_cgraph * gf, + ggml_tensor * wo, + ggml_tensor * wo_b, + ggml_tensor * q_cur, + ggml_tensor * k_cur, + ggml_tensor * v_cur, + ggml_tensor * kq_b, + float kq_scale, + int il) const; + + llm_graph_input_attn_cross * build_attn_inp_cross() const; + + ggml_tensor * build_attn( + llm_graph_input_attn_cross * inp, + ggml_cgraph * gf, + ggml_tensor * wo, + ggml_tensor * wo_b, + ggml_tensor * q_cur, + ggml_tensor * k_cur, + ggml_tensor * v_cur, + ggml_tensor * kq_b, + float kq_scale, + int il) const; + + // + // recurrent + // + + ggml_tensor * build_copy_mask_state( + ggml_cgraph * gf, + ggml_tensor * s, + ggml_tensor * state_copy, + ggml_tensor * state_mask, + int32_t n_state, + int32_t n_seqs) const; + + ggml_tensor * build_rwkv_token_shift_load( + ggml_cgraph * gf, + ggml_tensor * state_copy, + ggml_tensor * state_mask, + const llama_ubatch & ubatch, + int il) const; + + ggml_tensor * build_rwkv_token_shift_store( + ggml_tensor * token_shift, + const llama_ubatch & ubatch, + int il) const; + + // + // pooling + // + + void build_pooling( + ggml_cgraph * gf, + ggml_tensor * cls, + ggml_tensor * cls_b, + ggml_tensor * cls_out, + ggml_tensor * cls_out_b) const; +}; diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp index ea87b2953d..58e98bf231 100644 --- a/src/llama-hparams.cpp +++ b/src/llama-hparams.cpp @@ -69,3 +69,11 @@ uint32_t llama_hparams::n_embd_v_s() const { // corresponds to Mamba's ssm_states size return ssm_d_state * ssm_d_inner; } + +bool llama_hparams::is_sliding(uint32_t il) const { + if (il < n_layer) { + return n_swa > 0 && n_swa_pattern > 0 && il % n_swa_pattern < (n_swa_pattern - 1); + } + + GGML_ABORT("fatal error"); +} diff --git a/src/llama-hparams.h b/src/llama-hparams.h index 1fe4541037..e3091c8127 100644 --- a/src/llama-hparams.h +++ b/src/llama-hparams.h @@ -36,6 +36,7 @@ struct llama_hparams { uint32_t n_layer; uint32_t n_rot; uint32_t n_swa = 0; // sliding window attention (SWA) + uint32_t n_swa_pattern = 1; // by default, all layers use non-sliding-window attention uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head uint32_t n_expert = 0; @@ -133,6 +134,8 @@ struct llama_hparams { // dimension of the recurrent state embeddings uint32_t n_embd_v_s() const; + + bool is_sliding(uint32_t il) const; }; static_assert(std::is_trivially_copyable::value, "llama_hparams must be trivially copyable"); diff --git a/src/llama-io.cpp b/src/llama-io.cpp new file mode 100644 index 0000000000..7ad70d1633 --- /dev/null +++ b/src/llama-io.cpp @@ -0,0 +1,15 @@ +#include "llama-io.h" + +void llama_io_write_i::write_string(const std::string & str) { + uint32_t str_size = str.size(); + + write(&str_size, sizeof(str_size)); + write(str.data(), str_size); +} + +void llama_io_read_i::read_string(std::string & str) { + uint32_t str_size; + read_to(&str_size, sizeof(str_size)); + + str.assign((const char *) read(str_size), str_size); +} diff --git a/src/llama-io.h b/src/llama-io.h new file mode 100644 index 0000000000..ce9216b83b --- /dev/null +++ b/src/llama-io.h @@ -0,0 +1,35 @@ +#pragma once + +#include +#include +#include + +struct ggml_tensor; + +class llama_io_write_i { +public: + llama_io_write_i() = default; + virtual ~llama_io_write_i() = default; + + virtual void write(const void * src, size_t size) = 0; + virtual void write_tensor(const ggml_tensor * tensor, size_t offset, size_t size) = 0; + + // bytes written so far + virtual size_t n_bytes() = 0; + + void write_string(const std::string & str); +}; + +class llama_io_read_i { +public: + llama_io_read_i() = default; + virtual ~llama_io_read_i() = default; + + virtual const uint8_t * read(size_t size) = 0; + virtual void read_to(void * dst, size_t size) = 0; + + // bytes read so far + virtual size_t n_bytes() = 0; + + void read_string(std::string & str); +}; diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp index feffdf0de5..14c8933b4d 100644 --- a/src/llama-kv-cache.cpp +++ b/src/llama-kv-cache.cpp @@ -6,86 +6,92 @@ #include "llama-model.h" #include +#include #include #include +#include static const llama_kv_cache_slot_info llama_kv_cache_slot_info_failed{false}; -uint32_t llama_kv_cache_get_padding(const struct llama_cparams & cparams) { - // the FA kernels require padding to avoid extra runtime boundary checks - return cparams.flash_attn ? 256u : 32u; +llama_kv_cache_unified::llama_kv_cache_unified(const llama_hparams & hparams, callbacks cbs) : hparams(hparams), cbs(std::move(cbs)) { } -bool llama_kv_cache_init( - struct llama_kv_cache & cache, - const llama_model & model, - const llama_cparams & cparams, - ggml_type type_k, - ggml_type type_v, - uint32_t kv_size, - bool offload) { - const struct llama_hparams & hparams = model.hparams; - +bool llama_kv_cache_unified::init( + const llama_model & model, + const llama_cparams & cparams, + ggml_type type_k, + ggml_type type_v, + uint32_t kv_size, + bool offload) { const int32_t n_layer = hparams.n_layer; - cache.has_shift = false; + has_shift = false; - cache.recurrent = llama_model_is_recurrent(&model); - cache.v_trans = !cache.recurrent && !cparams.flash_attn; - cache.can_shift = !cache.recurrent && model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA + recurrent = llama_model_is_recurrent(&model); + v_trans = !recurrent && !cparams.flash_attn; + can_shift = !recurrent && model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA LLAMA_LOG_INFO("%s: kv_size = %d, offload = %d, type_k = '%s', type_v = '%s', n_layer = %d, can_shift = %d\n", - __func__, kv_size, offload, ggml_type_name(type_k), ggml_type_name(type_v), n_layer, cache.can_shift); + __func__, kv_size, offload, ggml_type_name(type_k), ggml_type_name(type_v), n_layer, can_shift); - cache.head = 0; - cache.size = kv_size; - cache.used = 0; + head = 0; + size = kv_size; + used = 0; - cache.type_k = type_k; - cache.type_v = type_v; + this->type_k = type_k; + this->type_v = type_v; - cache.cells.clear(); - cache.cells.resize(kv_size); + cells.clear(); + cells.resize(kv_size); // create a context for each buffer type std::map ctx_map; auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * { auto it = ctx_map.find(buft); if (it == ctx_map.end()) { - struct ggml_init_params params = { + ggml_init_params params = { /*.mem_size =*/ size_t(2u*n_layer*ggml_tensor_overhead()), /*.mem_buffer =*/ NULL, /*.no_alloc =*/ true, }; + ggml_context * ctx = ggml_init(params); if (!ctx) { return nullptr; } + ctx_map[buft] = ctx; - cache.ctxs.emplace_back(ctx); + ctxs.emplace_back(ctx); + return ctx; } + return it->second; }; - cache.k_l.reserve(n_layer); - cache.v_l.reserve(n_layer); + k_l.reserve(n_layer); + v_l.reserve(n_layer); for (int i = 0; i < n_layer; i++) { const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s(); const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s(); - LLAMA_LOG_DEBUG("%s: layer %d: n_embd_k_gqa = %d, n_embd_v_gqa = %d\n", __func__, i, n_embd_k_gqa, n_embd_v_gqa); + const char * dev_name = "CPU"; ggml_backend_buffer_type_t buft; if (offload) { auto * dev = model.dev_layer(i); buft = ggml_backend_dev_buffer_type(dev); + + dev_name = ggml_backend_dev_name(dev); } else { buft = ggml_backend_cpu_buffer_type(); } - ggml_context * ctx = ctx_for_buft(buft); + LLAMA_LOG_DEBUG("%s: layer %3d: n_embd_k_gqa = %d, n_embd_v_gqa = %d, dev = %s\n", __func__, + i, n_embd_k_gqa, n_embd_v_gqa, dev_name); + + ggml_context * ctx = ctx_for_buft(buft); if (!ctx) { LLAMA_LOG_ERROR("%s: failed to create ggml context for kv cache\n", __func__); return false; @@ -95,8 +101,8 @@ bool llama_kv_cache_init( ggml_tensor * v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size); ggml_format_name(k, "cache_k_l%d", i); ggml_format_name(v, "cache_v_l%d", i); - cache.k_l.push_back(k); - cache.v_l.push_back(v); + k_l.push_back(k); + v_l.push_back(v); } // allocate tensors and initialize the buffers to avoid NaNs in the padding @@ -111,280 +117,80 @@ bool llama_kv_cache_init( } ggml_backend_buffer_clear(buf, 0); LLAMA_LOG_INFO("%s: %10s KV buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0); - cache.bufs.emplace_back(buf); + bufs.emplace_back(buf); } return true; } -struct llama_kv_cache_slot_info llama_kv_cache_find_slot( - struct llama_kv_cache & cache, - const struct llama_ubatch & ubatch) { - const uint32_t n_tokens = ubatch.n_tokens; - const uint32_t n_seqs = ubatch.n_seqs; - const uint32_t n_seq_tokens = ubatch.n_seq_tokens; +int32_t llama_kv_cache_unified::get_n_tokens() const { + int32_t result = 0; - if (cache.recurrent) { - // For recurrent state architectures (like Mamba or RWKV), - // each cache cell can store the state for a whole sequence. - // A slot should be always be contiguous. - - // can only process batches with an equal number of new tokens in each sequence - GGML_ASSERT(ubatch.equal_seqs); - - int32_t min = cache.size - 1; - int32_t max = 0; - - // everything should fit if all seq_ids are smaller than the max - for (uint32_t s = 0; s < n_seqs; ++s) { - const uint32_t n_seq_id = ubatch.n_seq_id[s]; - for (uint32_t j = 0; j < n_seq_id; ++j) { - const llama_seq_id seq_id = ubatch.seq_id[s][j]; - - if (seq_id < 0 || (uint32_t) seq_id >= cache.size) { - // too big seq_id - // TODO: would it be possible to resize the cache instead? - LLAMA_LOG_ERROR("%s: seq_id=%d >= n_seq_max=%d Try using a bigger --parallel value\n", __func__, seq_id, cache.size); - return llama_kv_cache_slot_info_failed; - } - if (j > 0) { - llama_kv_cell & seq = cache.cells[seq_id]; - if (seq.tail >= 0) { - llama_kv_cell & cell = cache.cells[seq.tail]; - // clear cells from seq_ids that become shared - // (should not normally happen, but let's handle it anyway) - cell.seq_id.erase(seq_id); - seq.tail = -1; - if (cell.seq_id.empty()) { - cell.pos = -1; - cell.src = -1; - cache.used -= 1; - } - } - } - } - } - -#ifndef NDEBUG - { - std::vector tails_verif; - tails_verif.assign(cache.size, -1); - for (uint32_t i = 0; i < cache.size; ++i) { - llama_kv_cell & cell = cache.cells[i]; - for (llama_seq_id seq_id : cell.seq_id) { - if (tails_verif[seq_id] != -1) { - LLAMA_LOG_ERROR("%s: duplicate tail for seq_id %d in cell %d and %d\n", __func__, seq_id, i, tails_verif[seq_id]); - } - tails_verif[seq_id] = i; - } - } - for (uint32_t i = 0; i < cache.size; ++i) { - if (tails_verif[i] != cache.cells[i].tail) { - LLAMA_LOG_ERROR("%s: wrong tail for seq_id %d, (%d instead of %d)\n", __func__, i, cache.cells[i].tail, tails_verif[i]); - } - } - } -#endif - - // find next empty cell - uint32_t next_empty_cell = cache.head; - - for (uint32_t i = 0; i < cache.size; ++i) { - if (next_empty_cell >= cache.size) { next_empty_cell -= cache.size; } - llama_kv_cell & cell = cache.cells[next_empty_cell]; - if (cell.is_empty()) { break; } - next_empty_cell += 1; - } - - // find usable cell range - for (uint32_t s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; - llama_kv_cell & seq_meta = cache.cells[seq_id]; - bool has_cell = false; - if (seq_meta.tail >= 0) { - llama_kv_cell & cell = cache.cells[seq_meta.tail]; - GGML_ASSERT(cell.has_seq_id(seq_id)); - // does this seq_id "own" the cell? - if (cell.seq_id.size() == 1) { has_cell = true; } - } - if (!has_cell) { - llama_kv_cell & empty_cell = cache.cells[next_empty_cell]; - GGML_ASSERT(empty_cell.is_empty()); - // copy old tail into the empty cell - if (seq_meta.tail >= 0) { - llama_kv_cell & orig_cell = cache.cells[seq_meta.tail]; - empty_cell.pos = orig_cell.pos; - empty_cell.src = orig_cell.src; - orig_cell.seq_id.erase(seq_id); - empty_cell.seq_id.insert(seq_id); // will be overwritten - } - seq_meta.tail = next_empty_cell; - // find next empty cell - if (s + 1 < n_seqs) { - next_empty_cell += 1; - for (uint32_t i = 0; i < cache.size; ++i) { - if (next_empty_cell >= cache.size) { next_empty_cell -= cache.size; } - llama_kv_cell & cell = cache.cells[next_empty_cell]; - if (cell.is_empty()) { break; } - next_empty_cell += 1; - } - } - } - if (min > seq_meta.tail) { min = seq_meta.tail; } - if (max < seq_meta.tail) { max = seq_meta.tail; } - } - - // gather and re-order - for (uint32_t s = 0; s < n_seqs; ++s) { - int32_t dst_id = s + min; - int32_t src_id = cache.cells[ubatch.seq_id[s][0]].tail; - if (dst_id != src_id) { - llama_kv_cell & dst_cell = cache.cells[dst_id]; - llama_kv_cell & src_cell = cache.cells[src_id]; - - std::swap(dst_cell.pos, src_cell.pos); - std::swap(dst_cell.src, src_cell.src); - std::swap(dst_cell.seq_id, src_cell.seq_id); - - // swap tails (assuming they NEVER overlap) - for (const llama_seq_id seq_id : src_cell.seq_id) { - cache.cells[seq_id].tail = src_id; - } - for (const llama_seq_id seq_id : dst_cell.seq_id) { - cache.cells[seq_id].tail = dst_id; - } - } - } - - // update the pos of the used seqs - for (uint32_t s = 0; s < n_seqs; ++s) { - const llama_pos last_pos = ubatch.pos[n_seq_tokens * s + n_seq_tokens - 1]; - int32_t cell_id = s + min; - llama_kv_cell & cell = cache.cells[cell_id]; - - if (cell.pos >= 0 && last_pos != cell.pos + (llama_pos) n_seq_tokens) { - // What should happen when the pos backtracks or skips a value? - // Clearing the state mid-batch would require special-casing which isn't done. - LLAMA_LOG_WARN("%s: non-consecutive token position %d after %d for sequence %d with %u new tokens\n", - __func__, last_pos, cell.pos, ubatch.seq_id[s][0], n_seq_tokens); - } - cell.pos = last_pos; - cell.seq_id.clear(); - for (int32_t j = 0; j < ubatch.n_seq_id[s]; ++j) { - const llama_seq_id seq_id = ubatch.seq_id[s][j]; - cell.seq_id.insert(seq_id); - cache.cells[seq_id].tail = cell_id; - } - } - - // allow getting the range of used cells, from head to head + n - cache.head = min; - cache.n = max - min + 1; - cache.used = std::count_if(cache.cells.begin(), cache.cells.end(), - [](const llama_kv_cell& cell){ return !cell.is_empty(); }); - - // sanity check - return llama_kv_cache_slot_info(cache.n >= n_seqs); - } - // otherwise, one cell per token. - - if (n_tokens > cache.size) { - LLAMA_LOG_ERROR("%s: n_tokens=%d > cache.size=%d\n", __func__, n_tokens, cache.size); - return llama_kv_cache_slot_info_failed; + for (uint32_t i = 0; i < size; i++) { + result += cells[i].seq_id.size(); } - uint32_t n_tested = 0; - - while (true) { - if (cache.head + n_tokens > cache.size) { - n_tested += cache.size - cache.head; - cache.head = 0; - continue; - } - - bool found = true; - for (uint32_t i = 0; i < n_tokens; i++) { - if (cache.cells[cache.head + i].pos >= 0) { - found = false; - cache.head += i + 1; - n_tested += i + 1; - break; - } - } - - if (found) { - break; - } - - if (n_tested >= cache.size) { - //LLAMA_LOG_ERROR("%s: failed to find a slot for %d tokens\n", __func__, n_tokens); - return llama_kv_cache_slot_info_failed; - } - } - - for (uint32_t s = 0; s < n_seqs; s++) { - for (uint32_t i = 0; i < n_seq_tokens; ++i) { - uint32_t k = s*n_seq_tokens + i; - cache.cells[cache.head + k].pos = ubatch.pos[k]; - - for (int32_t j = 0; j < ubatch.n_seq_id[s]; j++) { - cache.cells[cache.head + k].seq_id.insert(ubatch.seq_id[s][j]); - } - } - } - - cache.used += n_tokens; - - return llama_kv_cache_slot_info(cache.head, cache.head + n_tokens); + return result; } -uint32_t llama_kv_cache_cell_max(const struct llama_kv_cache & cache) { - for (uint32_t i = cache.size; i > 0; --i) { - const llama_kv_cell & cell = cache.cells[i - 1]; - - if (cell.pos >= 0 && !cell.is_empty()) { - return i; - } - } - - return 0; +uint32_t llama_kv_cache_unified::get_used_cells() const { + return used; } -void llama_kv_cache_clear(struct llama_kv_cache & cache) { - for (int32_t i = 0; i < (int32_t) cache.size; ++i) { - cache.cells[i].pos = -1; - cache.cells[i].seq_id.clear(); - cache.cells[i].src = -1; - cache.cells[i].tail = -1; +size_t llama_kv_cache_unified::total_size() const { + size_t size = 0; + for (const auto & buf : bufs) { + size += ggml_backend_buffer_get_size(buf.get()); } - cache.head = 0; - cache.used = 0; - for (auto & buf : cache.bufs) { + return size; +} + +llama_pos llama_kv_cache_unified::pos_max() const { + llama_pos pos_max = -1; + for (const auto & cell : cells) { + pos_max = std::max(pos_max, cell.pos); + } + + return pos_max; +} + +void llama_kv_cache_unified::clear() { + for (int32_t i = 0; i < (int32_t) size; ++i) { + cells[i].pos = -1; + cells[i].seq_id.clear(); + cells[i].src = -1; + cells[i].tail = -1; + } + head = 0; + used = 0; + + for (auto & buf : bufs) { ggml_backend_buffer_clear(buf.get(), 0); } } -bool llama_kv_cache_seq_rm( - struct llama_kv_cache & cache, - llama_seq_id seq_id, - llama_pos p0, - llama_pos p1) { - uint32_t new_head = cache.size; +bool llama_kv_cache_unified::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) { + uint32_t new_head = size; - if (p0 < 0) p0 = 0; - if (p1 < 0) p1 = std::numeric_limits::max(); + if (p0 < 0) { + p0 = 0; + } + + if (p1 < 0) { + p1 = std::numeric_limits::max(); + } // models like Mamba or RWKV can't have a state partially erased - if (cache.recurrent) { - if (seq_id >= (int64_t) cache.size) { + if (recurrent) { + if (seq_id >= (int64_t) size) { // could be fatal return false; } if (0 <= seq_id) { - int32_t & tail_id = cache.cells[seq_id].tail; + int32_t & tail_id = cells[seq_id].tail; if (tail_id >= 0) { - const llama_kv_cell & cell = cache.cells[tail_id]; + const llama_kv_cell & cell = cells[tail_id]; // partial intersection is invalid if ((0 < p0 && p0 <= cell.pos) || (0 < p1 && p1 <= cell.pos)) { return false; @@ -402,48 +208,59 @@ bool llama_kv_cache_seq_rm( } } - for (uint32_t i = 0; i < cache.size; ++i) { - if (cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { + for (uint32_t i = 0; i < size; ++i) { + if (cells[i].pos >= p0 && cells[i].pos < p1) { if (seq_id < 0) { - cache.cells[i].seq_id.clear(); - } else if (cache.cells[i].has_seq_id(seq_id)) { - cache.cells[i].seq_id.erase(seq_id); + cells[i].seq_id.clear(); + } else if (cells[i].has_seq_id(seq_id)) { + cells[i].seq_id.erase(seq_id); } else { continue; } - if (cache.cells[i].is_empty()) { + if (cells[i].is_empty()) { // keep count of the number of used cells - if (cache.cells[i].pos >= 0) cache.used--; + if (cells[i].pos >= 0) { + used--; + } - cache.cells[i].pos = -1; - cache.cells[i].src = -1; - if (new_head == cache.size) new_head = i; + cells[i].pos = -1; + cells[i].src = -1; + + if (new_head == size) { + new_head = i; + } } } } // If we freed up a slot, set head to it so searching can start there. - if (new_head != cache.size && new_head < cache.head) cache.head = new_head; + if (new_head != size && new_head < head) { + head = new_head; + } return true; } -void llama_kv_cache_seq_cp( - struct llama_kv_cache & cache, - llama_seq_id seq_id_src, - llama_seq_id seq_id_dst, - llama_pos p0, - llama_pos p1) { - if (p0 < 0) p0 = 0; - if (p1 < 0) p1 = std::numeric_limits::max(); +void llama_kv_cache_unified::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) { + if (seq_id_src == seq_id_dst) { + return; + } - if (cache.recurrent) { - if ((uint32_t) seq_id_dst < cache.size && (uint32_t) seq_id_src < cache.size) { - llama_kv_cell & tail_src = cache.cells[seq_id_src]; - llama_kv_cell & tail_dst = cache.cells[seq_id_dst]; + if (p0 < 0) { + p0 = 0; + } + + if (p1 < 0) { + p1 = std::numeric_limits::max(); + } + + if (recurrent) { + if ((uint32_t) seq_id_dst < size && (uint32_t) seq_id_src < size) { + llama_kv_cell & tail_src = cells[seq_id_src]; + llama_kv_cell & tail_dst = cells[seq_id_dst]; if (tail_dst.tail >= 0) { // clear destination seq_id if it wasn't empty - llama_kv_cell & cell_dst = cache.cells[tail_dst.tail]; + llama_kv_cell & cell_dst = cells[tail_dst.tail]; cell_dst.seq_id.erase(seq_id_dst); tail_dst.tail = -1; @@ -451,11 +268,11 @@ void llama_kv_cache_seq_cp( cell_dst.pos = -1; cell_dst.delta = -1; cell_dst.src = -1; - cache.used -= 1; + used -= 1; } } if (tail_src.tail >= 0) { - llama_kv_cell & cell_src = cache.cells[tail_src.tail]; + llama_kv_cell & cell_src = cells[tail_src.tail]; cell_src.seq_id.insert(seq_id_dst); tail_dst.tail = tail_src.tail; @@ -464,59 +281,75 @@ void llama_kv_cache_seq_cp( return; } - // otherwise, this is the KV cache of a Transformer-like model - cache.head = 0; + // otherwise, this is the KV of a Transformer-like model + head = 0; - for (uint32_t i = 0; i < cache.size; ++i) { - if (cache.cells[i].has_seq_id(seq_id_src) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { - cache.cells[i].seq_id.insert(seq_id_dst); + for (uint32_t i = 0; i < size; ++i) { + if (cells[i].has_seq_id(seq_id_src) && cells[i].pos >= p0 && cells[i].pos < p1) { + cells[i].seq_id.insert(seq_id_dst); } } } -void llama_kv_cache_seq_keep(struct llama_kv_cache & cache, llama_seq_id seq_id) { - uint32_t new_head = cache.size; +void llama_kv_cache_unified::seq_keep(llama_seq_id seq_id) { + uint32_t new_head = size; - for (uint32_t i = 0; i < cache.size; ++i) { - if (cache.recurrent && (llama_seq_id) i != seq_id) { - cache.cells[i].tail = -1; + for (uint32_t i = 0; i < size; ++i) { + if (recurrent && (llama_seq_id) i != seq_id) { + cells[i].tail = -1; } - if (!cache.cells[i].has_seq_id(seq_id)) { - if (cache.cells[i].pos >= 0) cache.used--; - cache.cells[i].pos = -1; - cache.cells[i].src = -1; - cache.cells[i].seq_id.clear(); - if (new_head == cache.size) new_head = i; + + if (!cells[i].has_seq_id(seq_id)) { + if (cells[i].pos >= 0) { + used--; + } + + cells[i].pos = -1; + cells[i].src = -1; + cells[i].seq_id.clear(); + + if (new_head == size){ + new_head = i; + } } else { - cache.cells[i].seq_id.clear(); - cache.cells[i].seq_id.insert(seq_id); + cells[i].seq_id.clear(); + cells[i].seq_id.insert(seq_id); } } // If we freed up a slot, set head to it so searching can start there. - if (new_head != cache.size && new_head < cache.head) cache.head = new_head; + if (new_head != size && new_head < head) { + head = new_head; + } } -void llama_kv_cache_seq_add( - struct llama_kv_cache & cache, - llama_seq_id seq_id, - llama_pos p0, - llama_pos p1, - llama_pos delta) { - uint32_t new_head = cache.size; +void llama_kv_cache_unified::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) { + if (delta == 0) { + return; + } - if (p0 < 0) p0 = 0; - if (p1 < 0) p1 = std::numeric_limits::max(); - // If there is no range then return early to avoid looping over the cache. - if (p0 == p1) return; + uint32_t new_head = size; - if (cache.recurrent) { + if (p0 < 0) { + p0 = 0; + } + + if (p1 < 0) { + p1 = std::numeric_limits::max(); + } + + // If there is no range then return early to avoid looping over the + if (p0 == p1) { + return; + } + + if (recurrent) { // for Mamba-like or RWKV models, only the pos needs to be shifted - if (0 <= seq_id && seq_id < (int64_t) cache.size) { - const int32_t tail_id = cache.cells[seq_id].tail; + if (0 <= seq_id && seq_id < (int64_t) size) { + const int32_t tail_id = cells[seq_id].tail; if (tail_id >= 0) { - llama_kv_cell & cell = cache.cells[tail_id]; + llama_kv_cell & cell = cells[tail_id]; if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) { cell.pos += delta; } @@ -525,19 +358,19 @@ void llama_kv_cache_seq_add( return; } - for (uint32_t i = 0; i < cache.size; ++i) { - if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { - cache.has_shift = true; - cache.cells[i].pos += delta; - cache.cells[i].delta += delta; + for (uint32_t i = 0; i < size; ++i) { + if (cells[i].has_seq_id(seq_id) && cells[i].pos >= p0 && cells[i].pos < p1) { + has_shift = true; + cells[i].pos += delta; + cells[i].delta += delta; - if (cache.cells[i].pos < 0) { - if (!cache.cells[i].is_empty()) { - cache.used--; + if (cells[i].pos < 0) { + if (!cells[i].is_empty()) { + used--; } - cache.cells[i].pos = -1; - cache.cells[i].seq_id.clear(); - if (new_head == cache.size) { + cells[i].pos = -1; + cells[i].seq_id.clear(); + if (new_head == size) { new_head = i; } } @@ -546,93 +379,968 @@ void llama_kv_cache_seq_add( // If we freed up a slot, set head to it so searching can start there. // Otherwise we just start the next search from the beginning. - cache.head = new_head != cache.size ? new_head : 0; + head = new_head != size ? new_head : 0; } -void llama_kv_cache_seq_div( - struct llama_kv_cache & cache, - llama_seq_id seq_id, - llama_pos p0, - llama_pos p1, - int d) { - if (p0 < 0) p0 = 0; - if (p1 < 0) p1 = std::numeric_limits::max(); - // If there is no range then return early to avoid looping over the cache. - if (p0 == p1) return; +void llama_kv_cache_unified::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) { + if (d == 1) { + return; + } - if (cache.recurrent) { + if (p0 < 0) { + p0 = 0; + } + + if (p1 < 0) { + p1 = std::numeric_limits::max(); + } + + // If there is no range then return early to avoid looping over the cache. + if (p0 == p1) { + return; + } + + if (recurrent) { // for Mamba-like or RWKV models, only the pos needs to be changed - if (0 <= seq_id && seq_id < (int64_t) cache.size) { - const int32_t tail_id = cache.cells[seq_id].tail; + if (0 <= seq_id && seq_id < (int64_t) size) { + const int32_t tail_id = cells[seq_id].tail; if (tail_id >= 0) { - llama_kv_cell & cell = cache.cells[tail_id]; + llama_kv_cell & cell = cells[tail_id]; if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) { cell.pos /= d; } } } + return; } - for (uint32_t i = 0; i < cache.size; ++i) { - if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { - cache.has_shift = true; + for (uint32_t i = 0; i < size; ++i) { + if (cells[i].has_seq_id(seq_id) && cells[i].pos >= p0 && cells[i].pos < p1) { + has_shift = true; { - llama_pos p_old = cache.cells[i].pos; - cache.cells[i].pos /= d; - cache.cells[i].delta += cache.cells[i].pos - p_old; + llama_pos p_old = cells[i].pos; + cells[i].pos /= d; + cells[i].delta += cells[i].pos - p_old; } } } } -llama_pos llama_kv_cache_seq_pos_max(struct llama_kv_cache & cache, llama_seq_id seq_id) { +llama_pos llama_kv_cache_unified::seq_pos_max(llama_seq_id seq_id) { llama_pos result = 0; - for (uint32_t i = 0; i < cache.size; ++i) { - if (cache.cells[i].has_seq_id(seq_id)) { - result = std::max(result, cache.cells[i].pos); + for (uint32_t i = 0; i < size; ++i) { + if (cells[i].has_seq_id(seq_id)) { + result = std::max(result, cells[i].pos); } } return result; } -void llama_kv_cache_defrag(struct llama_kv_cache & cache) { - if (!cache.recurrent) { - cache.do_defrag = true; +void llama_kv_cache_unified::defrag() { + if (!recurrent) { + do_defrag = true; } } -int32_t llama_get_kv_cache_token_count(const struct llama_kv_cache & kv) { - int result = 0; +bool llama_kv_cache_unified::get_can_shift() const { + return can_shift; +} - for (uint32_t i = 0; i < kv.size; i++) { - result += kv.cells[i].seq_id.size(); +llama_kv_cache_slot_info llama_kv_cache_unified::find_slot( + const llama_ubatch & ubatch) { + const uint32_t n_tokens = ubatch.n_tokens; + const uint32_t n_seqs = ubatch.n_seqs; + const uint32_t n_seq_tokens = ubatch.n_seq_tokens; + + if (recurrent) { + // For recurrent state architectures (like Mamba or RWKV), + // each cache cell can store the state for a whole sequence. + // A slot should be always be contiguous. + + // can only process batches with an equal number of new tokens in each sequence + GGML_ASSERT(ubatch.equal_seqs); + + int32_t min = size - 1; + int32_t max = 0; + + // everything should fit if all seq_ids are smaller than the max + for (uint32_t s = 0; s < n_seqs; ++s) { + const uint32_t n_seq_id = ubatch.n_seq_id[s]; + for (uint32_t j = 0; j < n_seq_id; ++j) { + const llama_seq_id seq_id = ubatch.seq_id[s][j]; + + if (seq_id < 0 || (uint32_t) seq_id >= size) { + // too big seq_id + // TODO: would it be possible to resize the cache instead? + LLAMA_LOG_ERROR("%s: seq_id=%d >= n_seq_max=%d Try using a bigger --parallel value\n", __func__, seq_id, size); + return llama_kv_cache_slot_info_failed; + } + if (j > 0) { + llama_kv_cell & seq = cells[seq_id]; + if (seq.tail >= 0) { + llama_kv_cell & cell = cells[seq.tail]; + // clear cells from seq_ids that become shared + // (should not normally happen, but let's handle it anyway) + cell.seq_id.erase(seq_id); + seq.tail = -1; + if (cell.seq_id.empty()) { + cell.pos = -1; + cell.src = -1; + used -= 1; + } + } + } + } + } + +#ifndef NDEBUG + { + std::vector tails_verif; + tails_verif.assign(size, -1); + for (uint32_t i = 0; i < size; ++i) { + llama_kv_cell & cell = cells[i]; + for (llama_seq_id seq_id : cell.seq_id) { + if (tails_verif[seq_id] != -1) { + LLAMA_LOG_ERROR("%s: duplicate tail for seq_id %d in cell %d and %d\n", __func__, seq_id, i, tails_verif[seq_id]); + } + tails_verif[seq_id] = i; + } + } + for (uint32_t i = 0; i < size; ++i) { + if (tails_verif[i] != cells[i].tail) { + LLAMA_LOG_ERROR("%s: wrong tail for seq_id %d, (%d instead of %d)\n", __func__, i, cells[i].tail, tails_verif[i]); + } + } + } +#endif + + // find next empty cell + uint32_t next_empty_cell = head; + + for (uint32_t i = 0; i < size; ++i) { + if (next_empty_cell >= size) { next_empty_cell -= size; } + llama_kv_cell & cell = cells[next_empty_cell]; + if (cell.is_empty()) { break; } + next_empty_cell += 1; + } + + // find usable cell range + for (uint32_t s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = ubatch.seq_id[s][0]; + llama_kv_cell & seq_meta = cells[seq_id]; + bool has_cell = false; + if (seq_meta.tail >= 0) { + llama_kv_cell & cell = cells[seq_meta.tail]; + GGML_ASSERT(cell.has_seq_id(seq_id)); + // does this seq_id "own" the cell? + if (cell.seq_id.size() == 1) { has_cell = true; } + } + if (!has_cell) { + llama_kv_cell & empty_cell = cells[next_empty_cell]; + GGML_ASSERT(empty_cell.is_empty()); + // copy old tail into the empty cell + if (seq_meta.tail >= 0) { + llama_kv_cell & orig_cell = cells[seq_meta.tail]; + empty_cell.pos = orig_cell.pos; + empty_cell.src = orig_cell.src; + orig_cell.seq_id.erase(seq_id); + empty_cell.seq_id.insert(seq_id); // will be overwritten + } + seq_meta.tail = next_empty_cell; + // find next empty cell + if (s + 1 < n_seqs) { + next_empty_cell += 1; + for (uint32_t i = 0; i < size; ++i) { + if (next_empty_cell >= size) { next_empty_cell -= size; } + llama_kv_cell & cell = cells[next_empty_cell]; + if (cell.is_empty()) { break; } + next_empty_cell += 1; + } + } + } + if (min > seq_meta.tail) { min = seq_meta.tail; } + if (max < seq_meta.tail) { max = seq_meta.tail; } + } + + // gather and re-order + for (uint32_t s = 0; s < n_seqs; ++s) { + int32_t dst_id = s + min; + int32_t src_id = cells[ubatch.seq_id[s][0]].tail; + if (dst_id != src_id) { + llama_kv_cell & dst_cell = cells[dst_id]; + llama_kv_cell & src_cell = cells[src_id]; + + std::swap(dst_cell.pos, src_cell.pos); + std::swap(dst_cell.src, src_cell.src); + std::swap(dst_cell.seq_id, src_cell.seq_id); + + // swap tails (assuming they NEVER overlap) + for (const llama_seq_id seq_id : src_cell.seq_id) { + cells[seq_id].tail = src_id; + } + for (const llama_seq_id seq_id : dst_cell.seq_id) { + cells[seq_id].tail = dst_id; + } + } + } + + // update the pos of the used seqs + for (uint32_t s = 0; s < n_seqs; ++s) { + const llama_pos last_pos = ubatch.pos[n_seq_tokens * s + n_seq_tokens - 1]; + int32_t cell_id = s + min; + llama_kv_cell & cell = cells[cell_id]; + + if (cell.pos >= 0 && last_pos != cell.pos + (llama_pos) n_seq_tokens) { + // What should happen when the pos backtracks or skips a value? + // Clearing the state mid-batch would require special-casing which isn't done. + LLAMA_LOG_WARN("%s: non-consecutive token position %d after %d for sequence %d with %u new tokens\n", + __func__, last_pos, cell.pos, ubatch.seq_id[s][0], n_seq_tokens); + } + cell.pos = last_pos; + cell.seq_id.clear(); + for (int32_t j = 0; j < ubatch.n_seq_id[s]; ++j) { + const llama_seq_id seq_id = ubatch.seq_id[s][j]; + cell.seq_id.insert(seq_id); + cells[seq_id].tail = cell_id; + } + } + + // allow getting the range of used cells, from head to head + n + head = min; + n = max - min + 1; + used = std::count_if(cells.begin(), cells.end(), + [](const llama_kv_cell& cell){ return !cell.is_empty(); }); + + // sanity check + return llama_kv_cache_slot_info(n >= n_seqs); } - return result; + // otherwise, one cell per token. + + if (n_tokens > size) { + LLAMA_LOG_ERROR("%s: n_tokens = %d > size = %d\n", __func__, n_tokens, size); + return llama_kv_cache_slot_info_failed; + } + + uint32_t n_tested = 0; + + while (true) { + if (head + n_tokens > size) { + n_tested += size - head; + head = 0; + continue; + } + + bool found = true; + for (uint32_t i = 0; i < n_tokens; i++) { + if (cells[head + i].pos >= 0) { + found = false; + head += i + 1; + n_tested += i + 1; + break; + } + } + + if (found) { + break; + } + + if (n_tested >= size) { + //LLAMA_LOG_ERROR("%s: failed to find a slot for %d tokens\n", __func__, n_tokens); + return llama_kv_cache_slot_info_failed; + } + } + + for (uint32_t s = 0; s < n_seqs; s++) { + for (uint32_t i = 0; i < n_seq_tokens; ++i) { + uint32_t k = s*n_seq_tokens + i; + cells[head + k].pos = ubatch.pos[k]; + + for (int32_t j = 0; j < ubatch.n_seq_id[s]; j++) { + cells[head + k].seq_id.insert(ubatch.seq_id[s][j]); + } + } + } + + used += n_tokens; + + return llama_kv_cache_slot_info(head, head + n_tokens); } -int32_t llama_get_kv_cache_used_cells(const struct llama_kv_cache & kv) { - return kv.used; +uint32_t llama_kv_cache_unified::get_padding(const llama_cparams & cparams) const { + // the FA kernels require padding to avoid extra runtime boundary checks + return cparams.flash_attn ? 256u : 32u; } -bool llama_kv_cache_can_shift(const struct llama_kv_cache & kv) { - return kv.can_shift; +uint32_t llama_kv_cache_unified::cell_max() const { + for (uint32_t i = size; i > 0; --i) { + const llama_kv_cell & cell = cells[i - 1]; + + if (cell.pos >= 0 && !cell.is_empty()) { + return i; + } + } + + return 0; +} + +size_t llama_kv_cache_unified::size_k_bytes() const { + size_t size_k_bytes = 0; + + for (const auto & k : k_l) { + size_k_bytes += ggml_nbytes(k); + } + + return size_k_bytes; +} + +size_t llama_kv_cache_unified::size_v_bytes() const { + size_t size_v_bytes = 0; + + for (const auto & v : v_l) { + size_v_bytes += ggml_nbytes(v); + } + + return size_v_bytes; +} + +bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) { + const uint32_t n_layer = hparams.n_layer; + + const uint32_t n_kv = cell_max(); + const uint32_t n_used = used; + + assert(n_used <= n_kv); + + //const int64_t t_start = ggml_time_us(); + + // number of cells moved + uint32_t n_moves = 0; + + // each move requires 6*n_layer tensors (see graph_build_kv_self_defrag) + // - source view, destination view, copy operation + // - x2 for keys and values + //const uint32_t max_moves = max_nodes()/(6*n_layer); + // TODO: tmp fix https://github.com/ggerganov/llama.cpp/issues/6685#issuecomment-2057579516 + const uint32_t max_moves = (n_max_nodes - 2*n_layer)/(6*n_layer); + + // determine which KV cells to move where + // + // cell i moves to ids[i] + // + // if ids[i] == i || ids[i] == n_kv, then cell i is not moved + // + auto & ids = defrag_info.ids; + + ids.clear(); + ids.resize(n_kv, n_kv); + + for (uint32_t i0 = 0; i0 < n_used; ++i0) { + const auto & cell0 = cells[i0]; + + if (!cell0.is_empty()) { + ids[i0] = i0; + + continue; + } + + // found a hole - fill it with data from the end of the cache + + uint32_t nh = 1; + + // determine the size of the hole + while (i0 + nh < n_used && cells[i0 + nh].is_empty()) { + nh++; + } + + uint32_t nf = 0; + uint32_t is = n_kv - 1; + + // starting from the end, find nh non-empty cells + for (; is > i0; --is) { + const auto & cell1 = cells[is]; + + if (cell1.is_empty() || ids[is] != n_kv) { + continue; + } + + // non-empty cell which is not yet moved + nf++; + + if (nf == nh) { + break; + } + } + + // this can only happen if `n_used` is not accurate, which would be a bug + GGML_ASSERT(nf == nh && "KV defrag bug: nf != nh"); + + nf = 0; + + uint32_t i1 = is; + + // are we moving a continuous block of memory? + bool cont = false; + + // should we stop searching for the next move? + bool stop = false; + + // go back and move the nf cells to the hole + for (; i1 < n_kv; ++i1) { + auto & cell1 = cells[i1]; + + if (cell1.is_empty() || ids[i1] != n_kv) { + if (n_moves == max_moves) { + stop = true; + break; + } + + cont = false; + continue; + } + + // this cell goes to (i0 + nf) + ids[i1] = i0 + nf; + + // move the cell meta data + cells[i0 + nf] = cell1; + + // clear the old cell and move the head there + cell1 = llama_kv_cell(); + head = n_used; + + if (!cont) { + n_moves++; + cont = true; + } + + nf++; + + if (nf == nh) { + break; + } + } + + if (stop || n_moves == max_moves) { + break; + } + + //LLAMA_LOG_INFO("(tmp log) KV defrag: move [%u, %u) to [%u, %u)\n", is, i1 + 1, i0, i0 + nh); + + i0 += nh - 1; + } + + if (n_moves == 0) { + return false; + } + + LLAMA_LOG_DEBUG("(tmp log) KV defrag cell moves: %u\n", n_moves); + + LLAMA_LOG_DEBUG("expected gf nodes: %u\n", 6*n_moves*n_layer); + + return true; +} + +void llama_kv_cache_unified::state_write(llama_io_write_i & io, llama_seq_id seq_id) const { + std::vector> cell_ranges; // ranges, from inclusive, to exclusive + uint32_t cell_count = 0; + + // Count the number of cells with the specified seq_id + // Find all the ranges of cells with this seq id (or all, when -1) + uint32_t cell_range_begin = size; + for (uint32_t i = 0; i < size; ++i) { + const auto & cell = cells[i]; + if ((seq_id == -1 && !cell.is_empty()) || cell.has_seq_id(seq_id)) { + ++cell_count; + if (cell_range_begin == size) { + cell_range_begin = i; + } + } else { + if (cell_range_begin != size) { + cell_ranges.emplace_back(cell_range_begin, i); + cell_range_begin = size; + } + } + } + if (cell_range_begin != size) { + cell_ranges.emplace_back(cell_range_begin, size); + } + + // DEBUG CHECK: Sum of cell counts in ranges should equal the total cell count + uint32_t cell_count_check = 0; + for (const auto & range : cell_ranges) { + cell_count_check += range.second - range.first; + } + GGML_ASSERT(cell_count == cell_count_check); + + io.write(&cell_count, sizeof(cell_count)); + + state_write_meta(io, cell_ranges, seq_id); + state_write_data(io, cell_ranges); +} + +void llama_kv_cache_unified::state_read(llama_io_read_i & io, llama_seq_id seq_id) { + uint32_t cell_count; + io.read_to(&cell_count, sizeof(cell_count)); + + bool res = true; + res = res && state_read_meta(io, cell_count, seq_id); + res = res && state_read_data(io, cell_count); + + if (!res) { + if (seq_id == -1) { + clear(); + } else { + seq_rm(seq_id, -1, -1); + } + throw std::runtime_error("failed to restore kv cache"); + } +} + +void llama_kv_cache_unified::state_write_meta(llama_io_write_i & io, const std::vector> & cell_ranges, llama_seq_id seq_id) const { + for (const auto & range : cell_ranges) { + for (uint32_t i = range.first; i < range.second; ++i) { + const auto & cell = cells[i]; + const llama_pos pos = cell.pos; + const uint32_t n_seq_id = seq_id == -1 ? cell.seq_id.size() : 0; + + io.write(&pos, sizeof(pos)); + io.write(&n_seq_id, sizeof(n_seq_id)); + + if (n_seq_id) { + for (auto seq_id : cell.seq_id) { + io.write(&seq_id, sizeof(seq_id)); + } + } + } + } +} + +void llama_kv_cache_unified::state_write_data(llama_io_write_i & io, const std::vector> & cell_ranges) const { + const uint32_t v_trans = this->v_trans ? 1 : 0; + const uint32_t n_layer = hparams.n_layer; + + io.write(&v_trans, sizeof(v_trans)); + io.write(&n_layer, sizeof(n_layer)); + + std::vector tmp_buf; + + // Iterate and write all the keys first, each row is a cell + // Get whole range at a time + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); + + // Write key type + const int32_t k_type_i = (int32_t)k_l[il]->type; + io.write(&k_type_i, sizeof(k_type_i)); + + // Write row size of key + const uint64_t k_size_row = ggml_row_size(k_l[il]->type, n_embd_k_gqa); + io.write(&k_size_row, sizeof(k_size_row)); + + // Read each range of cells of k_size length each into tmp_buf and write out + for (const auto & range : cell_ranges) { + const size_t range_size = range.second - range.first; + const size_t buf_size = range_size * k_size_row; + io.write_tensor(k_l[il], range.first * k_size_row, buf_size); + } + } + + if (!v_trans) { + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + + // Write value type + const int32_t v_type_i = (int32_t)v_l[il]->type; + io.write(&v_type_i, sizeof(v_type_i)); + + // Write row size of value + const uint64_t v_size_row = ggml_row_size(v_l[il]->type, n_embd_v_gqa); + io.write(&v_size_row, sizeof(v_size_row)); + + // Read each range of cells of v_size length each into tmp_buf and write out + for (const auto & range : cell_ranges) { + const size_t range_size = range.second - range.first; + const size_t buf_size = range_size * v_size_row; + io.write_tensor(v_l[il], range.first * v_size_row, buf_size); + } + } + } else { + // When v is transposed, we also need the element size and get the element ranges from each row + const uint32_t kv_size = size; + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + + // Write value type + const int32_t v_type_i = (int32_t)v_l[il]->type; + io.write(&v_type_i, sizeof(v_type_i)); + + // Write element size + const uint32_t v_size_el = ggml_type_size(v_l[il]->type); + io.write(&v_size_el, sizeof(v_size_el)); + + // Write GQA embedding size + io.write(&n_embd_v_gqa, sizeof(n_embd_v_gqa)); + + // For each row, we get the element values of each cell + for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { + // Read each range of cells of v_size_el length each into tmp_buf and write out + for (const auto & range : cell_ranges) { + const size_t range_size = range.second - range.first; + const size_t src_offset = (range.first + j * kv_size) * v_size_el; + const size_t buf_size = range_size * v_size_el; + io.write_tensor(v_l[il], src_offset, buf_size); + } + } + } + } +} + +bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id) { + if (dest_seq_id != -1) { + // single sequence + + seq_rm(dest_seq_id, -1, -1); + + llama_sbatch sbatch; + llama_ubatch batch = sbatch.reserve_ubatch(cell_count, /* has_embd */ false); + + batch.n_tokens = cell_count; + batch.n_seq_tokens = cell_count; + batch.n_seqs = 1; + + for (uint32_t i = 0; i < cell_count; ++i) { + llama_pos pos; + uint32_t n_seq_id; + + io.read_to(&pos, sizeof(pos)); + io.read_to(&n_seq_id, sizeof(n_seq_id)); + + if (n_seq_id != 0) { + LLAMA_LOG_ERROR("%s: invalid seq_id-agnostic kv cell\n", __func__); + return false; + } + + batch.pos[i] = pos; + } + batch.n_seq_id[0] = 1; + batch.seq_id[0] = &dest_seq_id; + if (!find_slot(batch)) { + LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__); + return false; + } + + // DEBUG CHECK: kv.head should be our first cell, kv.head + cell_count - 1 should be our last cell (verify seq_id and pos values) + // Assume that this is one contiguous block of cells + GGML_ASSERT(head + cell_count <= size); + GGML_ASSERT(cells[head].pos == batch.pos[0]); + GGML_ASSERT(cells[head + cell_count - 1].pos == batch.pos[cell_count - 1]); + GGML_ASSERT(cells[head].has_seq_id(dest_seq_id)); + GGML_ASSERT(cells[head + cell_count - 1].has_seq_id(dest_seq_id)); + } else { + // whole KV cache restore + + if (cell_count > size) { + LLAMA_LOG_ERROR("%s: not enough cells in kv cache\n", __func__); + return false; + } + + clear(); + + for (uint32_t i = 0; i < cell_count; ++i) { + llama_kv_cell & cell = cells[i]; + + llama_pos pos; + uint32_t n_seq_id; + + io.read_to(&pos, sizeof(pos)); + io.read_to(&n_seq_id, sizeof(n_seq_id)); + + cell.pos = pos; + + for (uint32_t j = 0; j < n_seq_id; ++j) { + llama_seq_id seq_id; + io.read_to(&seq_id, sizeof(seq_id)); + + // TODO: llama_kv_cache_unified should have a notion of max sequences + //if (seq_id < 0 || (uint32_t) seq_id >= llama_n_seq_max(ctx)) { + if (seq_id < 0) { + //LLAMA_LOG_ERROR("%s: invalid seq_id, %d is out of range [0, %u)\n", __func__, seq_id, llama_n_seq_max(ctx)); + LLAMA_LOG_ERROR("%s: invalid seq_id, %d is out of range [0, inf)\n", __func__, seq_id); + return false; + } + + cell.seq_id.insert(seq_id); + + if (recurrent) { + int32_t & tail = cells[seq_id].tail; + if (tail != -1) { + LLAMA_LOG_ERROR("%s: duplicate tail for seq_id %d in cell %d and %d\n", __func__, seq_id, i, tail); + return false; + } + tail = i; + } + } + } + + head = 0; + used = cell_count; + } + + if (recurrent) { + for (uint32_t i = 0; i < cell_count; ++i) { + uint32_t cell_id = head + i; + // make sure the recurrent states will keep their restored state + cells[cell_id].src = cell_id; + } + } + + return true; +} + +bool llama_kv_cache_unified::state_read_data(llama_io_read_i & io, uint32_t cell_count) { + uint32_t v_trans; + uint32_t n_layer; + io.read_to(&v_trans, sizeof(v_trans)); + io.read_to(&n_layer, sizeof(n_layer)); + + if (n_layer != hparams.n_layer) { + LLAMA_LOG_ERROR("%s: mismatched layer count (%u instead of %u)\n", __func__, n_layer, hparams.n_layer); + return false; + } + if (cell_count > size) { + LLAMA_LOG_ERROR("%s: not enough cells in kv cache to restore state (%u > %u)\n", __func__, cell_count, size); + return false; + } + if (v_trans != (bool) v_trans) { + LLAMA_LOG_ERROR("%s: incompatible V transposition\n", __func__); + return false; + } + + // For each layer, read the keys for each cell, one row is one cell, read as one contiguous block + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); + + // Read type of key + int32_t k_type_i_ref; + io.read_to(&k_type_i_ref, sizeof(k_type_i_ref)); + const int32_t k_type_i = (int32_t) k_l[il]->type; + if (k_type_i != k_type_i_ref) { + LLAMA_LOG_ERROR("%s: mismatched key type (%d != %d, layer %d)\n", __func__, k_type_i, k_type_i_ref, il); + return false; + } + + // Read row size of key + uint64_t k_size_row_ref; + io.read_to(&k_size_row_ref, sizeof(k_size_row_ref)); + const size_t k_size_row = ggml_row_size(k_l[il]->type, n_embd_k_gqa); + if (k_size_row != k_size_row_ref) { + LLAMA_LOG_ERROR("%s: mismatched key row size (%zu != %zu, layer %d)\n", __func__, k_size_row, (size_t) k_size_row_ref, il); + return false; + } + + if (cell_count) { + // Read and set the keys for the whole cell range + ggml_backend_tensor_set(k_l[il], io.read(cell_count * k_size_row), head * k_size_row, cell_count * k_size_row); + } + } + + if (!v_trans) { + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + + // Read type of value + int32_t v_type_i_ref; + io.read_to(&v_type_i_ref, sizeof(v_type_i_ref)); + const int32_t v_type_i = (int32_t)v_l[il]->type; + if (v_type_i != v_type_i_ref) { + LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il); + return false; + } + + // Read row size of value + uint64_t v_size_row_ref; + io.read_to(&v_size_row_ref, sizeof(v_size_row_ref)); + const size_t v_size_row = ggml_row_size(v_l[il]->type, n_embd_v_gqa); + if (v_size_row != v_size_row_ref) { + LLAMA_LOG_ERROR("%s: mismatched value row size (%zu != %zu, layer %d)\n", __func__, v_size_row, (size_t) v_size_row_ref, il); + return false; + } + + if (cell_count) { + // Read and set the values for the whole cell range + ggml_backend_tensor_set(v_l[il], io.read(cell_count * v_size_row), head * v_size_row, cell_count * v_size_row); + } + } + } else { + // For each layer, read the values for each cell (transposed) + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + + // Read type of value + int32_t v_type_i_ref; + io.read_to(&v_type_i_ref, sizeof(v_type_i_ref)); + const int32_t v_type_i = (int32_t)v_l[il]->type; + if (v_type_i != v_type_i_ref) { + LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il); + return false; + } + + // Read element size of value + uint32_t v_size_el_ref; + io.read_to(&v_size_el_ref, sizeof(v_size_el_ref)); + const size_t v_size_el = ggml_type_size(v_l[il]->type); + if (v_size_el != v_size_el_ref) { + LLAMA_LOG_ERROR("%s: mismatched value element size (%zu != %zu, layer %d)\n", __func__, v_size_el, (size_t) v_size_el_ref, il); + return false; + } + + // Read GQA embedding size + uint32_t n_embd_v_gqa_ref; + io.read_to(&n_embd_v_gqa_ref, sizeof(n_embd_v_gqa_ref)); + if (n_embd_v_gqa != n_embd_v_gqa_ref) { + LLAMA_LOG_ERROR("%s: mismatched GQA embedding size (%u != %u, layer %d)\n", __func__, n_embd_v_gqa, n_embd_v_gqa_ref, il); + return false; + } + + if (cell_count) { + // For each row in the transposed matrix, read the values for the whole cell range + for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { + const size_t dst_offset = (head + j * size) * v_size_el; + ggml_backend_tensor_set(v_l[il], io.read(cell_count * v_size_el), dst_offset, cell_count * v_size_el); + } + } + } + } + + return true; +} + +// +// interface implementation +// + +int32_t llama_kv_cache_n_tokens(const llama_kv_cache * kv) { + if (!kv) { + return 0; + } + + return kv->get_n_tokens(); +} + +int32_t llama_kv_cache_used_cells(const llama_kv_cache * kv) { + if (!kv) { + return 0; + } + + return kv->get_used_cells(); +} + +void llama_kv_cache_clear(llama_kv_cache * kv) { + if (!kv) { + return; + } + + kv->clear(); +} + +bool llama_kv_cache_seq_rm( + llama_kv_cache * kv, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1) { + if (!kv) { + return true; + } + + return kv->seq_rm(seq_id, p0, p1); +} + +void llama_kv_cache_seq_cp( + llama_kv_cache * kv, + llama_seq_id seq_id_src, + llama_seq_id seq_id_dst, + llama_pos p0, + llama_pos p1) { + if (!kv) { + return; + } + + kv->seq_cp(seq_id_src, seq_id_dst, p0, p1); +} + +void llama_kv_cache_seq_keep(llama_kv_cache * kv, llama_seq_id seq_id) { + if (!kv) { + return; + } + + kv->seq_keep(seq_id); +} + +void llama_kv_cache_seq_add( + llama_kv_cache * kv, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + llama_pos delta) { + if (!kv) { + return; + } + + kv->seq_add(seq_id, p0, p1, delta); +} + +void llama_kv_cache_seq_div( + llama_kv_cache * kv, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + int d) { + if (!kv) { + return; + } + + kv->seq_div(seq_id, p0, p1, d); +} + +llama_pos llama_kv_cache_seq_pos_max(llama_kv_cache * kv, llama_seq_id seq_id) { + if (!kv) { + return 0; + } + + return kv->seq_pos_max(seq_id); +} + +void llama_kv_cache_defrag(llama_kv_cache * kv) { + if (!kv) { + return; + } + + kv->defrag(); +} + +bool llama_kv_cache_can_shift(const llama_kv_cache * kv) { + if (!kv) { + return false; + } + + return kv->get_can_shift(); } // // kv cache view // -struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_kv_cache & kv, int32_t n_seq_max) { - struct llama_kv_cache_view result = { +llama_kv_cache_view llama_kv_cache_view_init(const llama_kv_cache & kv, int32_t n_seq_max) { + llama_kv_cache_view result = { /*.n_cells = */ 0, /*.n_seq_max = */ n_seq_max, /*.token_count = */ 0, - /*.used_cells = */ llama_get_kv_cache_used_cells(kv), + /*.used_cells = */ llama_kv_cache_used_cells(&kv), /*.max_contiguous = */ 0, /*.max_contiguous_idx = */ -1, /*.cells = */ nullptr, @@ -642,7 +1350,7 @@ struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_kv_cache return result; } -void llama_kv_cache_view_free(struct llama_kv_cache_view * view) { +void llama_kv_cache_view_free(llama_kv_cache_view * view) { if (view->cells != nullptr) { free(view->cells); view->cells = nullptr; @@ -653,18 +1361,25 @@ void llama_kv_cache_view_free(struct llama_kv_cache_view * view) { } } -void llama_kv_cache_view_update(struct llama_kv_cache_view * view, const struct llama_kv_cache & kv) { - if (uint32_t(view->n_cells) < kv.size || view->cells == nullptr) { - view->n_cells = int32_t(kv.size); - void * p = realloc(view->cells, sizeof(struct llama_kv_cache_view_cell) * view->n_cells); +void llama_kv_cache_view_update(llama_kv_cache_view * view, const llama_kv_cache * kv) { + // TODO: rework this in the future, for now quick hack + const llama_kv_cache_unified * kvu = dynamic_cast(kv); + if (kvu == nullptr) { + LLAMA_LOG_ERROR("%s: the kv_cache_view currently works only with llama_kv_cache_unified\n", __func__); + return; + } + + if (uint32_t(view->n_cells) < kvu->size || view->cells == nullptr) { + view->n_cells = int32_t(kvu->size); + void * p = realloc(view->cells, sizeof(llama_kv_cache_view_cell) * view->n_cells); GGML_ASSERT(p != nullptr && "Failed to alloc kv_cache_view cells"); - view->cells = (struct llama_kv_cache_view_cell *)p; + view->cells = (llama_kv_cache_view_cell *)p; p = realloc(view->cells_sequences, sizeof(llama_seq_id) * view->n_seq_max * view->n_cells); GGML_ASSERT(p != nullptr && "Failed to alloc kv_cache_view cells sequences"); view->cells_sequences = (llama_seq_id *)p; } - const std::vector & kv_cells = kv.cells; + const std::vector & kv_cells = kvu->cells; llama_kv_cache_view_cell * c_curr = view->cells; llama_seq_id * cs_curr = view->cells_sequences; int32_t used_cells = 0; @@ -673,7 +1388,7 @@ void llama_kv_cache_view_update(struct llama_kv_cache_view * view, const struct uint32_t max_contig = 0; int32_t max_contig_idx = -1; - for (int32_t i = 0; i < int32_t(kv.size); i++, c_curr++, cs_curr += view->n_seq_max) { + for (int32_t i = 0; i < int32_t(kvu->size); i++, c_curr++, cs_curr += view->n_seq_max) { const size_t curr_size = kv_cells[i].seq_id.size(); token_count += curr_size; c_curr->pos = kv_cells[i].pos + kv_cells[i].delta; @@ -711,8 +1426,8 @@ void llama_kv_cache_view_update(struct llama_kv_cache_view * view, const struct view->max_contiguous_idx = max_contig_idx; view->token_count = token_count; view->used_cells = used_cells; - if (uint32_t(used_cells) != kv.used) { + if (uint32_t(used_cells) != kvu->used) { LLAMA_LOG_ERROR("%s: used cells mismatch. kv_cache says %d but we calculated %d\n", - __func__, kv.used, used_cells); + __func__, kvu->used, used_cells); } } diff --git a/src/llama-kv-cache.h b/src/llama-kv-cache.h index 1ce0850ec8..0a7ff8a4ea 100644 --- a/src/llama-kv-cache.h +++ b/src/llama-kv-cache.h @@ -1,12 +1,29 @@ #pragma once #include "llama.h" +#include "llama-io.h" +#include "llama-memory.h" #include "ggml-cpp.h" +#include #include #include -#include + +struct llama_cparams; +struct llama_hparams; +struct llama_ubatch; + +struct llama_kv_cache : public llama_memory_i { + using llama_memory_i::llama_memory_i; + + virtual int32_t get_n_tokens() const = 0; + virtual uint32_t get_used_cells() const = 0; // TODO: remove, this is too-specific to the unified cache + + virtual bool get_can_shift() const = 0; + + bool get_can_edit() const override { return get_can_shift(); } +}; struct llama_kv_cell { llama_pos pos = -1; @@ -29,11 +46,105 @@ struct llama_kv_cell { } }; +// a structure holds information about the slot found in llama_kv_cache_find_slot +struct llama_kv_cache_slot_info { + std::pair boundaries; // slot boundaries [begin, end) + bool found = false; // the slot was found + + explicit llama_kv_cache_slot_info(bool found_) : found{found_} {} + llama_kv_cache_slot_info(uint32_t begin, uint32_t end) : boundaries{begin, end}, found{true} {} + + operator bool() const { return found; } +}; + // ring-buffer of cached KV data -struct llama_kv_cache { +// TODO: pimpl +// TODO: add notion of max sequences +class llama_kv_cache_unified : public llama_kv_cache { +public: + // can be used to query data from the model if needed + struct callbacks { + std::function get_rope_factors; + }; + + llama_kv_cache_unified( + const llama_hparams & hparams, + callbacks cbs); + + virtual ~llama_kv_cache_unified() = default; + + // TODO: become constructor + bool init( + const llama_model & model, // TODO: do not reference the model + const llama_cparams & cparams, + ggml_type type_k, + ggml_type type_v, + uint32_t kv_size, + bool offload); + + int32_t get_n_tokens() const override; + uint32_t get_used_cells() const override; + + size_t total_size() const; + + // TODO: better data structures to reduce the cost of this operation + llama_pos pos_max() const; + + void clear() override; + void defrag() override; + + bool seq_rm (llama_seq_id seq_id, llama_pos p0, llama_pos p1) override; + void seq_cp (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override; + void seq_keep(llama_seq_id seq_id) override; + void seq_add (llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) override; + void seq_div (llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) override; + + llama_pos seq_pos_max(llama_seq_id seq_id) override; + + bool get_can_shift() const override; + + // find an empty slot of size "n_tokens" in the cache + // updates the cache head + // returns a structure holding information about the slot found + // Note: On success, it's important that cache.head points + // to the first cell of the slot. + llama_kv_cache_slot_info find_slot(const llama_ubatch & batch); + + // TODO: maybe not needed + uint32_t get_padding(const llama_cparams & cparams) const; + + // find how many cells are currently in use + uint32_t cell_max() const; + + size_t size_k_bytes() const; + size_t size_v_bytes() const; + + // defrag + + struct { + std::vector ids; + } defrag_info; + + // return true if cells have been moved + bool defrag_prepare(int32_t n_max_nodes); + + // state save/load + + void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const; + void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1); + + // members + + const llama_hparams & hparams; + + callbacks cbs; + bool has_shift = false; bool do_defrag = false; + + // TODO: remove this and implement llama_kv_cache_recurrent instead bool recurrent = false; // with recurrent state models, a cell can hold the state for more than one past token + bool v_trans = true; // the value tensor is transposed bool can_shift = false; @@ -47,124 +158,30 @@ struct llama_kv_cache { // computed before each graph build uint32_t n = 0; + std::vector cells; + + std::vector k_l; // per layer + std::vector v_l; + +private: ggml_type type_k = GGML_TYPE_F16; ggml_type type_v = GGML_TYPE_F16; - std::vector cells; - - std::vector k_l; // per layer - std::vector v_l; - - std::vector ctxs; + std::vector ctxs; std::vector bufs; - size_t total_size() const { - size_t size = 0; - for (const auto & buf : bufs) { - size += ggml_backend_buffer_get_size(buf.get()); - } + void state_write_meta(llama_io_write_i & io, const std::vector> & cell_ranges, llama_seq_id seq_id = -1) const; + void state_write_data(llama_io_write_i & io, const std::vector> & cell_ranges) const; - return size; - } - - // TODO: better data structures to reduce the cost of this operation - llama_pos max_pos() const { - llama_pos max_pos = -1; - for (const auto & cell : cells) { - max_pos = std::max(max_pos, cell.pos); - } - - return max_pos; - } + bool state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id = -1); + bool state_read_data(llama_io_read_i & io, uint32_t cell_count); }; -// a structure holds information about the slot found in llama_kv_cache_find_slot -struct llama_kv_cache_slot_info { - std::pair boundaries; // slot boundaries [begin, end) - bool found = false; // the slot was found - - explicit llama_kv_cache_slot_info(bool found_) : found{found_} {} - llama_kv_cache_slot_info(uint32_t begin, uint32_t end) : boundaries{begin, end}, found{true} {} - - operator bool() const { return found; } -}; - -// TODO: maybe not needed -uint32_t llama_kv_cache_get_padding(const struct llama_cparams & cparams); - -bool llama_kv_cache_init( - struct llama_kv_cache & cache, - const llama_model & model, - const llama_cparams & cparams, - ggml_type type_k, - ggml_type type_v, - uint32_t kv_size, - bool offload); - -// find an empty slot of size "n_tokens" in the cache -// updates the cache head -// returns a structure holding information about the slot found -// Note: On success, it's important that cache.head points -// to the first cell of the slot. -struct llama_kv_cache_slot_info llama_kv_cache_find_slot( - struct llama_kv_cache & cache, - const struct llama_ubatch & batch); - -// find how many cells are currently in use -uint32_t llama_kv_cache_cell_max(const struct llama_kv_cache & cache); - -void llama_kv_cache_clear(struct llama_kv_cache & cache); - -bool llama_kv_cache_seq_rm( - struct llama_kv_cache & cache, - llama_seq_id seq_id, - llama_pos p0, - llama_pos p1); - -void llama_kv_cache_seq_cp( - struct llama_kv_cache & cache, - llama_seq_id seq_id_src, - llama_seq_id seq_id_dst, - llama_pos p0, - llama_pos p1); - -void llama_kv_cache_seq_keep( - struct llama_kv_cache & cache, - llama_seq_id seq_id); - -void llama_kv_cache_seq_add( - struct llama_kv_cache & cache, - llama_seq_id seq_id, - llama_pos p0, - llama_pos p1, - llama_pos delta); - -void llama_kv_cache_seq_div( - struct llama_kv_cache & cache, - llama_seq_id seq_id, - llama_pos p0, - llama_pos p1, - int d); - -llama_pos llama_kv_cache_seq_pos_max( - struct llama_kv_cache & cache, - llama_seq_id seq_id); - -void llama_kv_cache_defrag(struct llama_kv_cache & cache); - -int32_t llama_get_kv_cache_token_count(const struct llama_kv_cache & kv); - -int32_t llama_get_kv_cache_used_cells(const struct llama_kv_cache & kv); - -bool llama_kv_cache_can_shift(const struct llama_kv_cache & kv); - -// -// kv cache view -// - -struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_kv_cache & kv, int32_t n_seq_max); - -void llama_kv_cache_view_update(struct llama_kv_cache_view * view, const struct llama_kv_cache & kv); +// TODO: temporary reusing llama_kv_cache_unified -- implement recurrent cache and simplify llama_kv_cache_unified +//class llama_kv_cache_recurrent : public llama_kv_cache_unified { +//public: +// using llama_kv_cache_unified::llama_kv_cache_unified; +//}; // // kv cache restore @@ -184,13 +201,15 @@ struct llama_kv_slot_restorer { bool do_restore = false; - explicit llama_kv_slot_restorer(const struct llama_kv_cache & cache) { + llama_kv_cache_unified & cache; + + explicit llama_kv_slot_restorer(llama_kv_cache_unified & cache) : cache(cache) { old_state.head = cache.head; old_state.n = cache.n; } // saves a slot information for future restoration - void save(const struct llama_kv_cache_slot_info & slot) { + void save(const llama_kv_cache_slot_info & slot) { if (slot) { do_restore = true; if (slot.boundaries.first != slot.boundaries.second) { @@ -201,19 +220,68 @@ struct llama_kv_slot_restorer { // must be explicitly called to restore the kv_cache state // and rollback changes from all llama_kv_cache_find_slot calls - void restore(struct llama_kv_cache & cache) { + void restore() { if (do_restore) { cache.head = old_state.head; cache.n = old_state.n; if (cache.recurrent) { // recurrent models like Mamba or RWKV can't have a state partially erased - llama_kv_cache_seq_rm(cache, -1, -1, -1); + cache.seq_rm(-1, -1, -1); } else { for (auto & slot : slot_boundaries) { - llama_kv_cache_seq_rm(cache, -1, slot.first, slot.second); + cache.seq_rm(-1, slot.first, slot.second); } } } } }; +// TODO: maybe become part of the public llama_kv_cache in the future +int32_t llama_kv_cache_n_tokens(const llama_kv_cache * kv); + +int32_t llama_kv_cache_used_cells(const llama_kv_cache * kv); + +void llama_kv_cache_clear(llama_kv_cache * kv); + +bool llama_kv_cache_seq_rm( + llama_kv_cache * kv, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1); + +void llama_kv_cache_seq_cp( + llama_kv_cache * kv, + llama_seq_id seq_id_src, + llama_seq_id seq_id_dst, + llama_pos p0, + llama_pos p1); + +void llama_kv_cache_seq_keep(llama_kv_cache * kv, llama_seq_id seq_id); + +void llama_kv_cache_seq_add( + llama_kv_cache * kv, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + llama_pos delta); + +void llama_kv_cache_seq_div( + llama_kv_cache * kv, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + int d); + +llama_pos llama_kv_cache_seq_pos_max(llama_kv_cache * kv, llama_seq_id seq_id); + +void llama_kv_cache_defrag(llama_kv_cache * kv); + +bool llama_kv_cache_can_shift(const llama_kv_cache * kv); + +// +// kv cache view +// + +llama_kv_cache_view llama_kv_cache_view_init(const llama_kv_cache & kv, int32_t n_seq_max); + +void llama_kv_cache_view_update(llama_kv_cache_view * view, const llama_kv_cache * kv); diff --git a/src/llama-memory.cpp b/src/llama-memory.cpp new file mode 100644 index 0000000000..10173253ed --- /dev/null +++ b/src/llama-memory.cpp @@ -0,0 +1 @@ +#include "llama-memory.h" diff --git a/src/llama-memory.h b/src/llama-memory.h new file mode 100644 index 0000000000..69e6e34ca4 --- /dev/null +++ b/src/llama-memory.h @@ -0,0 +1,21 @@ +#pragma once + +#include "llama.h" + +// general concept of LLM memory +// the KV cache is a type of LLM memory, but there can be other types +class llama_memory_i { +public: + virtual void clear() = 0; + virtual void defrag() = 0; + + virtual bool seq_rm (llama_seq_id seq_id, llama_pos p0, llama_pos p1) = 0; + virtual void seq_cp (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) = 0; + virtual void seq_keep(llama_seq_id seq_id) = 0; + virtual void seq_add (llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) = 0; + virtual void seq_div (llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) = 0; + + virtual llama_pos seq_pos_max(llama_seq_id seq_id) = 0; + + virtual bool get_can_edit() const = 0; +}; diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 1da4eae7e6..5647d2ad62 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -2,13 +2,19 @@ #include "llama-impl.h" #include "llama-mmap.h" +#include "llama-batch.h" +#include "llama-cparams.h" #include "llama-model-loader.h" +#include "llama-kv-cache.h" #include "ggml-cpp.h" #include #include +#include +#include #include +#include #include #include #include @@ -243,6 +249,7 @@ static ggml_backend_buffer_type_t select_weight_buft(const llama_hparams & hpara return cur_buft; } } + return nullptr; } @@ -301,7 +308,7 @@ static buft_list_t make_cpu_buft_list(const std::vector & de } // GPU: split if LLAMA_SPLIT_MODE_ROW -> GPU -static buft_list_t make_gpu_buft_list(ggml_backend_dev_t dev, enum llama_split_mode split_mode, const float * tensor_split) { +static buft_list_t make_gpu_buft_list(ggml_backend_dev_t dev, llama_split_mode split_mode, const float * tensor_split) { buft_list_t buft_list; // add the device split buffer type if requested and available @@ -368,7 +375,7 @@ struct llama_model::impl { std::vector dev_layer; }; -llama_model::llama_model(const struct llama_model_params & params) : params(params), pimpl(std::make_unique()) { +llama_model::llama_model(const llama_model_params & params) : params(params), pimpl(std::make_unique()) { } llama_model::~llama_model() {} @@ -390,7 +397,7 @@ void llama_model::load_hparams(llama_model_loader & ml) { // get metadata as string for (int i = 0; i < gguf_get_n_kv(ctx); i++) { - enum gguf_type type = gguf_get_kv_type(ctx, i); + gguf_type type = gguf_get_kv_type(ctx, i); if (type == GGUF_TYPE_ARRAY) { continue; } @@ -851,11 +858,13 @@ void llama_model::load_hparams(llama_model_loader & ml) { case LLM_ARCH_GEMMA2: { hparams.n_swa = 4096; // default value of gemma 2 + hparams.n_swa_pattern = 2; + hparams.attn_soft_cap = true; + ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false); ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ml.get_key(LLM_KV_ATTN_LOGIT_SOFTCAPPING, hparams.f_attn_logit_softcapping, false); ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING, hparams.f_final_logit_softcapping, false); - hparams.attn_soft_cap = true; switch (hparams.n_layer) { case 26: type = LLM_TYPE_2B; break; @@ -864,6 +873,25 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_GEMMA3: + { + hparams.n_swa_pattern = 6; + + ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 26: type = LLM_TYPE_1B; break; + case 34: type = LLM_TYPE_4B; break; + case 48: type = LLM_TYPE_12B; break; + case 62: type = LLM_TYPE_27B; break; + default: type = LLM_TYPE_UNKNOWN; + } + + hparams.f_attention_scale = type == LLM_TYPE_27B + ? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0))) + : 1.0f / std::sqrt(float(hparams.n_embd_head_k)); + } break; case LLM_ARCH_STARCODER2: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); @@ -928,6 +956,8 @@ void llama_model::load_hparams(llama_model_loader & ml) { } break; case LLM_ARCH_COHERE2: { + hparams.n_swa_pattern = 4; + ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa); ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale); ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); @@ -1426,7 +1456,10 @@ bool llama_model::load_tensors(llama_model_loader & ml) { // skip unused tensors if (info.op == GGML_OP_NONE) { - LLAMA_LOG_WARN("model has unused tensor %s -- ignoring\n", tn.str().c_str()); + const size_t nbytes = ggml_nbytes(t_meta); + LLAMA_LOG_WARN("model has unused tensor %s (size = %zu bytes) -- ignoring\n", tn.str().c_str(), nbytes); + + ml.size_data -= nbytes; ml.n_created++; return nullptr; @@ -2454,6 +2487,35 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0); layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0); + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0); + layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0); + } + } break; + case LLM_ARCH_GEMMA3: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + // output + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0); + + layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0); + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0); + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0); + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); @@ -3584,8 +3646,8 @@ size_t llama_model::size() const { return pimpl->n_bytes; } -size_t llama_model::max_nodes() const { - return std::max(8192, tensors_by_name.size()*5); +size_t llama_model::n_tensors() const { + return tensors_by_name.size(); } size_t llama_model::n_devices() const { @@ -3650,6 +3712,7 @@ void llama_model::print_info() const { LLAMA_LOG_INFO("%s: f_clamp_kqv = %.1e\n", __func__, hparams.f_clamp_kqv); LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n", __func__, hparams.f_max_alibi_bias); LLAMA_LOG_INFO("%s: f_logit_scale = %.1e\n", __func__, hparams.f_logit_scale); + LLAMA_LOG_INFO("%s: f_attn_scale = %.1e\n", __func__, hparams.f_attention_scale); LLAMA_LOG_INFO("%s: n_ff = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_ff(il); }, hparams.n_layer).c_str()); LLAMA_LOG_INFO("%s: n_expert = %u\n", __func__, hparams.n_expert); LLAMA_LOG_INFO("%s: n_expert_used = %u\n", __func__, hparams.n_expert_used); @@ -3697,7 +3760,7 @@ void llama_model::print_info() const { LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared); LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale); LLAMA_LOG_INFO("%s: expert_weights_norm = %d\n", __func__, hparams.expert_weights_norm); - LLAMA_LOG_INFO("%s: expert_gating_func = %s\n", __func__, llama_expert_gating_func_name((enum llama_expert_gating_func_type) hparams.expert_gating_func)); + LLAMA_LOG_INFO("%s: expert_gating_func = %s\n", __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func)); LLAMA_LOG_INFO("%s: rope_yarn_log_mul = %.4f\n", __func__, hparams.rope_yarn_log_mul); } @@ -3773,9 +3836,9 @@ ggml_backend_buffer_type_t llama_model::select_buft(int il) const { }); } -const struct ggml_tensor * llama_model::get_tensor(const char * name) const { +const ggml_tensor * llama_model::get_tensor(const char * name) const { auto it = std::find_if(tensors_by_name.begin(), tensors_by_name.end(), - [name](const std::pair & it) { + [name](const std::pair & it) { return it.first == name; }); if (it == tensors_by_name.end()) { @@ -3785,12 +3848,7300 @@ const struct ggml_tensor * llama_model::get_tensor(const char * name) const { return it->second; } +struct llm_build_llama : public llm_graph_context { + llm_build_llama(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, kq_scale, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // For Granite architecture + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + if (model.layers[il].ffn_gate_inp == nullptr) { + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + } + + // For Granite architecture + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + // For Granite architecture + if (hparams.f_logit_scale) { + cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale); + } + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_deci : public llm_graph_context { + llm_build_deci(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + const int64_t n_head_kv = hparams.n_head_kv(il); + const int64_t n_head = hparams.n_head(il); + + if (n_head == 0) { + // attention-free layer of Llama-3_1-Nemotron-51B + cur = inpL; + } else { + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + } + + if (n_head > 0 && n_head_kv == 0) { + // "linear attention" of Llama-3_1-Nemotron-51B + cur = build_lora_mm(model.layers[il].wo, cur); + cb(cur, "wo", il); + } else if (n_head > 0) { + // self-attention + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, kq_scale, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // For Granite architecture + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + + // modified to support attention-free layer of Llama-3_1-Nemotron-51B + ggml_tensor * ffn_inp = cur; + if (n_head > 0) { + ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + } + + // feed-forward network + if (model.layers[il].ffn_gate_inp == nullptr) { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + // For Granite architecture + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + // For Granite architecture + if (hparams.f_logit_scale) { + cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale); + } + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_baichuan : public llm_graph_context { + llm_build_baichuan(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr; + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + switch (model.type) { + case LLM_TYPE_7B: + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + break; + case LLM_TYPE_13B: + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd/n_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd/n_head, n_head, n_tokens); + break; + default: + GGML_ABORT("fatal error"); + } + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_xverse : public llm_graph_context { + llm_build_xverse(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_falcon : public llm_graph_context { + llm_build_falcon(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * attn_norm; + + attn_norm = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(attn_norm, "attn_norm", il); + + // self-attention + { + if (model.layers[il].attn_norm_2) { + // Falcon-40B + cur = build_norm(inpL, + model.layers[il].attn_norm_2, + model.layers[il].attn_norm_2_b, + LLM_NORM, il); + cb(cur, "attn_norm_2", il); + } else { + cur = attn_norm; + } + + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + + // using mode = 2 for neox mode + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + attn_norm = ggml_get_rows(ctx0, attn_norm, inp_out_ids); + } + + ggml_tensor * ffn_inp = cur; + + // feed forward + { + cur = build_ffn(attn_norm, // !! use the attn norm, not the result + model.layers[il].ffn_up, NULL, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cur = ggml_add(ctx0, cur, inpL); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + // norm + cur = build_norm(cur, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_grok : public llm_graph_context { + llm_build_grok(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // multiply by embedding_multiplier_scale of 78.38367176906169 + inpL = ggml_scale(ctx0, inpL, 78.38367176906169f); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // Grok + // if attn_out_norm is present then apply it before adding the input + if (model.layers[il].attn_out_norm) { + cur = build_norm(cur, + model.layers[il].attn_out_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_out_norm", il); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_GELU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + + // Grok + // if layer_out_norm is present then apply it before adding the input + // Idea: maybe ffn_out_norm is a better name + if (model.layers[il].layer_out_norm) { + cur = build_norm(cur, + model.layers[il].layer_out_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "layer_out_norm", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + // Grok + // multiply logits by output_multiplier_scale of 0.5773502691896257 + + cur = ggml_scale(ctx0, cur, 0.5773502691896257f); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_dbrx : public llm_graph_context { + llm_build_dbrx(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = nullptr; + ggml_tensor * Kcur = nullptr; + ggml_tensor * Vcur = nullptr; + + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(cur, "wqkv_clamped", il); + + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].attn_out_norm, NULL, + LLM_NORM, il); + cb(cur, "attn_out_norm", il); + + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_starcoder : public llm_graph_context { + llm_build_starcoder(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + ggml_tensor * pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); + cb(pos, "pos_embd", -1); + + inpL = ggml_add(ctx0, inpL, pos); + cb(inpL, "inpL", -1); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_refact : public llm_graph_context { + llm_build_refact(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + cb(Kcur, "Kcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + cb(Qcur, "Qcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_bert : public llm_graph_context { + llm_build_bert(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + ggml_tensor * inp_pos = nullptr; + + if (model.arch != LLM_ARCH_JINA_BERT_V2) { + inp_pos = build_inp_pos(); + } + + // construct input embeddings (token, type, position) + inpL = build_inp_embd(model.tok_embd); + + // token types are hardcoded to zero ("Sentence A") + ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0); + inpL = ggml_add(ctx0, inpL, type_row0); + if (model.arch == LLM_ARCH_BERT) { + inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL); + } + cb(inpL, "inp_embd", -1); + + // embed layer norm + inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1); + cb(inpL, "inp_norm", -1); + + auto * inp_attn = build_attn_inp_no_cache(); + + // iterate layers + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * cur = inpL; + + ggml_tensor * Qcur; + ggml_tensor * Kcur; + ggml_tensor * Vcur; + + // self-attention + if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_JINA_BERT_V2) { + Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq); + + if (model.layers[il].attn_q_norm) { + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, + model.layers[il].attn_q_norm_b, + LLM_NORM, il); + } + + Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk); + + if (model.layers[il].attn_k_norm) { + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, + model.layers[il].attn_k_norm_b, + LLM_NORM, il); + } + + Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + } else { + // compute Q and K and RoPE them + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + cb(cur, "kqv_out", il); + + if (il == n_layer - 1 && pooling_type == LLAMA_POOLING_TYPE_NONE) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // re-add the layer input + cur = ggml_add(ctx0, cur, inpL); + + // attention layer norm + cur = build_norm(cur, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, il); + + if (model.layers[il].attn_norm_2 != nullptr) { + cur = ggml_add(ctx0, cur, inpL); // re-add the layer input + cur = build_norm(cur, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, il); + } + + ggml_tensor * ffn_inp = cur; + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + if (model.arch == LLM_ARCH_BERT) { + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + } else if (model.arch == LLM_ARCH_JINA_BERT_V2) { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_PAR, il); + } else { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + } + cb(cur, "ffn_out", il); + + // attentions bypass the intermediate layer + cur = ggml_add(ctx0, cur, ffn_inp); + + // output layer norm + cur = build_norm(cur, model.layers[il].layer_out_norm, model.layers[il].layer_out_norm_b, LLM_NORM, il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cb(cur, "result_embd", -1); + res->t_embd = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_bloom : public llm_graph_context { + llm_build_bloom(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + inpL = build_norm(inpL, + model.tok_norm, + model.tok_norm_b, + LLM_NORM, -1); + cb(inpL, "inp_norm", -1); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // Add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_mpt : public llm_graph_context { + llm_build_mpt(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * pos; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + if (model.pos_embd) { + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); + cb(pos, "pos_embd", -1); + + inpL = ggml_add(ctx0, inpL, pos); + cb(inpL, "inpL", -1); + } + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * attn_norm; + + attn_norm = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(attn_norm, "attn_norm", il); + + // self-attention + { + cur = attn_norm; + + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + if (model.layers[il].bqkv){ + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + } + + if (hparams.f_clamp_kqv > 0.0f) { + cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(cur, "wqkv_clamped", il); + } + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + // Q/K Layernorm + if (model.layers[il].attn_q_norm) { + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, + model.layers[il].attn_q_norm_b, + LLM_NORM, il); + cb(Qcur, "Qcur", il); + + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, + model.layers[il].attn_k_norm_b, + LLM_NORM, il); + cb(Kcur, "Kcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } else { + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // Add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // feed forward + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + model.layers[il].ffn_act, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_stablelm : public llm_graph_context { + llm_build_stablelm(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + ggml_tensor * inpSA = cur; + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + cb(Qcur, "Qcur", il); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + cb(Kcur, "Kcur", il); + + if (model.layers[il].attn_q_norm) { + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, + NULL, + LLM_NORM, il); + cb(Qcur, "Qcur", il); + } + if (model.layers[il].attn_k_norm) { + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, + NULL, + LLM_NORM, il); + cb(Kcur, "Kcur", il); + } + + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + if (model.layers[il].ffn_norm) { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + } else { + // parallel residual + cur = inpSA; + } + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_qwen : public llm_graph_context { + llm_build_qwen(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 2*sizeof(float)*(n_embd))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + + // using mode = 2 for neox mode + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward forward + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_qwen2 : public llm_graph_context { + llm_build_qwen2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_qwen2vl : public llm_graph_context { + llm_build_qwen2vl(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + int sections[4]; + std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + + Qcur = ggml_rope_multi( + ctx0, + ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_multi( + ctx0, + ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_qwen2moe : public llm_graph_context { + llm_build_qwen2moe(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * moe_out = + build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + + // FFN shared expert + { + ggml_tensor * cur_gate_inp = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur); + cb(cur_gate_inp, "ffn_shexp_gate_inp", il); + + // sigmoid + ggml_tensor * cur_gate = ggml_div(ctx0, ggml_silu(ctx0, cur_gate_inp), cur_gate_inp); + cb(cur_gate, "ffn_shexp_gate", il); + + ggml_tensor * cur_ffn = build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur_ffn, "ffn_shexp", il); + + ggml_tensor * ffn_shexp_out = ggml_mul(ctx0, cur_ffn, cur_gate); + cb(ffn_shexp_out, "ffn_shexp_out", il); + + moe_out = ggml_add(ctx0, moe_out, ffn_shexp_out); + cb(moe_out, "ffn_out", il); + + cur = moe_out; + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_phi2 : public llm_graph_context { + llm_build_phi2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * attn_norm_output; + ggml_tensor * ffn_output; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + attn_norm_output = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(attn_norm_output, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = nullptr; + ggml_tensor * Kcur = nullptr; + ggml_tensor * Vcur = nullptr; + + if (model.layers[il].wqkv) { + cur = build_lora_mm(model.layers[il].wqkv, attn_norm_output); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + } else { + Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq); + Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk); + Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + // with phi2, we scale the Q to avoid precision issues + // ref: https://github.com/ml-explore/mlx-examples/blob/08e862336ade809bc37d1035f94b359e7d1a5152/phi2/phi2.py#L64-L66 + Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head))); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids); + } + + // FF + { + ffn_output = build_ffn(attn_norm_output, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(ffn_output, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_output); + cur = ggml_add(ctx0, cur, inpL); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + cb(cur, "result_output_no_bias", -1); + + cur = ggml_add(ctx0, cur, model.output_b); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_phi3 : public llm_graph_context { + llm_build_phi3(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, true); + + for (int il = 0; il < n_layer; ++il) { + auto * residual = inpL; + + // self-attention + { + // rope freq factors for 128k context + ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + + ggml_tensor* attn_norm_output = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM_RMS, il); + cb(attn_norm_output, "attn_norm", il); + + ggml_tensor * Qcur = nullptr; + ggml_tensor * Kcur = nullptr; + ggml_tensor * Vcur = nullptr; + + if (model.layers[il].wqkv) { + cur = build_lora_mm(model.layers[il].wqkv, attn_norm_output); + cb(cur, "wqkv", il); + + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0 * sizeof(float) * (n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa))); + } else { + Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq); + Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk); + Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head))); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor* inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + residual = ggml_get_rows(ctx0, residual, inp_out_ids); + } + + cur = ggml_add(ctx0, cur, residual); + residual = cur; + + cur = build_norm(cur, + model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + if (model.layers[il].ffn_gate_inp == nullptr) { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SWIGLU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } else { + // MoE branch + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + } + + cur = ggml_add(ctx0, residual, cur); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + if (model.output_b != nullptr) { + cb(cur, "result_output_no_bias", -1); + cur = ggml_add(ctx0, cur, model.output_b); + } + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_plamo : public llm_graph_context { + llm_build_plamo(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + ggml_tensor * attention_norm = cur; + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_rot, n_head, n_tokens), inp_pos, nullptr, + n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens), inp_pos, nullptr, + n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + ggml_tensor * sa_out = cur; + + cur = attention_norm; + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + sa_out = ggml_get_rows(ctx0, sa_out, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // feed-forward network + { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, sa_out); + cur = ggml_add(ctx0, cur, inpL); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_gpt2 : public llm_graph_context { + llm_build_gpt2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * pos; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); + cb(pos, "pos_embd", -1); + + inpL = ggml_add(ctx0, inpL, pos); + cb(inpL, "inpL", -1); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_codeshell : public llm_graph_context { + llm_build_codeshell(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * tmpq = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + ggml_tensor * tmpk = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(tmpq, "tmpq", il); + cb(tmpk, "tmpk", il); + cb(Vcur, "Vcur", il); + + ggml_tensor * Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_orion : public llm_graph_context { + llm_build_orion(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + // if (model.layers[il].bq) { + // Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + // cb(Qcur, "Qcur", il); + // } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + // if (model.layers[il].bk) { + // Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + // cb(Kcur, "Kcur", il); + // } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + // if (model.layers[il].bv) { + // Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + // cb(Vcur, "Vcur", il); + // } + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_internlm2 : public llm_graph_context { + llm_build_internlm2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_minicpm3 : public llm_graph_context { + llm_build_minicpm3(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + //TODO: if the model varies, these parameters need to be read from the model + const int64_t n_embd_base = 256; + const float scale_embd = 12.0f; + const float scale_depth = 1.4f; + const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k)); + + const uint32_t n_embd_head_qk_rope = hparams.n_rot; + const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; + const uint32_t kv_lora_rank = hparams.n_lora_kv; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // scale the input embeddings + inpL = ggml_scale(ctx0, inpL, scale_embd); + cb(inpL, "inp_scaled", -1); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + ggml_tensor * q = NULL; + // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens} + q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur); + cb(q, "q", il); + + q = build_norm(q, + model.layers[il].attn_q_a_norm, NULL, + LLM_NORM_RMS, il); + cb(q, "q", il); + + // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens} + q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q); + cb(q, "q", il); + + // split into {n_head * n_embd_head_qk_nope, n_tokens} + ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + 0); + cb(q_nope, "q_nope", il); + + // and {n_head * n_embd_head_qk_rope, n_tokens} + ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + ggml_row_size(q->type, n_embd_head_qk_nope)); + cb(q_pe, "q_pe", il); + + // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens} + ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); + cb(kv_pe_compresseed, "kv_pe_compresseed", il); + + // split into {kv_lora_rank, n_tokens} + ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens, + kv_pe_compresseed->nb[1], + 0); + cb(kv_compressed, "kv_compressed", il); + + // and {n_embd_head_qk_rope, n_tokens} + ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens, + kv_pe_compresseed->nb[1], + kv_pe_compresseed->nb[1], + ggml_row_size(kv_pe_compresseed->type, kv_lora_rank)); + cb(k_pe, "k_pe", il); + + // TODO: the CUDA backend used to not support non-cont. (RMS) norm, investigate removing ggml_cont + kv_compressed = ggml_cont(ctx0, kv_compressed); + kv_compressed = build_norm(kv_compressed, + model.layers[il].attn_kv_a_norm, NULL, + LLM_NORM_RMS, il); + cb(kv_compressed, "kv_compressed", il); + + // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens} + ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed); + cb(kv, "kv", il); + + // split into {n_head * n_embd_head_qk_nope, n_tokens} + ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v), + ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)), + 0); + cb(k_nope, "k_nope", il); + + // and {n_head * n_embd_head_v, n_tokens} + ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens, + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)), + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head), + ggml_row_size(kv->type, (n_embd_head_qk_nope))); + cb(v_states, "v_states", il); + + v_states = ggml_cont(ctx0, v_states); + cb(v_states, "v_states", il); + + v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens, + ggml_row_size(kv->type, hparams.n_embd_head_v * n_head), + 0); + cb(v_states, "v_states", il); + + q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this + q_pe = ggml_rope_ext( + ctx0, q_pe, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(q_pe, "q_pe", il); + + // shared RoPE key + k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this + k_pe = ggml_rope_ext( + ctx0, k_pe, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(k_pe, "k_pe", il); + + ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0); + cb(q_states, "q_states", il); + + ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0); + cb(k_states, "k_states", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + q_states, k_states, v_states, nullptr, kq_scale, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // scale_res - scale the hidden states for residual connection + const float scale_res = scale_depth/sqrtf(float(n_layer)); + cur = ggml_scale(ctx0, cur, scale_res); + cb(cur, "hidden_scaled", il); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + // scale the hidden states for residual connection + cur = ggml_scale(ctx0, cur, scale_res); + cb(cur, "hidden_scaled_ffn", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head scaling + const float scale_lmhead = float(n_embd_base)/float(n_embd); + cur = ggml_scale(ctx0, cur, scale_lmhead); + cb(cur, "lmhead_scaling", -1); + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_gemma : public llm_graph_context { + llm_build_gemma(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head_k = hparams.n_embd_head_k; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); + cb(inpL, "inp_scaled", -1); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(Qcur, "Qcur", il); + + Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k))); + cb(Qcur, "Qcur_scaled", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); + cb(sa_out, "sa_out", il); + + cur = build_norm(sa_out, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, sa_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_gemma2 : public llm_graph_context { + llm_build_gemma2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head_k = hparams.n_embd_head_k; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); + cb(inpL, "inp_scaled", -1); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, true); + + for (int il = 0; il < n_layer; ++il) { + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(Qcur, "Qcur", il); + + // ref: https://github.com/google/gemma_pytorch/commit/03e657582d17cb5a8617ebf333c1c16f3694670e + switch (model.type) { + case LLM_TYPE_2B: + case LLM_TYPE_9B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k))); break; + case LLM_TYPE_27B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd / n_head))); break; + default: GGML_ABORT("fatal error"); + }; + cb(Qcur, "Qcur_scaled", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f, il); + } + + cur = build_norm(cur, + model.layers[il].attn_post_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); + cb(sa_out, "sa_out", il); + + cur = build_norm(sa_out, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = build_norm(cur, + model.layers[il].ffn_post_norm, NULL, + LLM_NORM_RMS, -1); + cb(cur, "ffn_post_norm", -1); + + cur = ggml_add(ctx0, cur, sa_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + // final logit soft-capping + cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping); + cur = ggml_tanh(ctx0, cur); + cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_gemma3 : public llm_graph_context { + llm_build_gemma3(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head_k = hparams.n_embd_head_k; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings) + if (ubatch.token) { + inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); + cb(inpL, "inp_scaled", -1); + } + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + // TODO: is causal == true correct? might need some changes + auto * inp_attn = build_attn_inp_kv_unified(true, true); + + for (int il = 0; il < n_layer; ++il) { + const bool is_sliding = hparams.is_sliding(il); + + const float freq_base_l = is_sliding ? 10000.0f : freq_base; + const float freq_scale_l = is_sliding ? 1.0f : freq_scale; + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head, n_tokens); + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(Qcur, "Qcur", il); + + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens); + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, hparams.f_attention_scale, il); + } + + cur = build_norm(cur, + model.layers[il].attn_post_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); + cb(sa_out, "sa_out", il); + + cur = build_norm(sa_out, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = build_norm(cur, + model.layers[il].ffn_post_norm, NULL, + LLM_NORM_RMS, -1); + cb(cur, "ffn_post_norm", -1); + + cur = ggml_add(ctx0, cur, sa_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +// TODO: move up next to build_starcoder +struct llm_build_starcoder2 : public llm_graph_context { + llm_build_starcoder2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_mamba : public llm_graph_context { + const llama_model & model; + + llm_build_mamba(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params), model(model) { + ggml_tensor * cur; + ggml_tensor * inpL; + + // {n_embd, n_tokens} + inpL = build_inp_embd(model.tok_embd); + + ggml_tensor * state_copy = build_inp_s_copy(); + ggml_tensor * state_mask = build_inp_s_mask(); + + for (int il = 0; il < n_layer; ++il) { + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + //cur = build_mamba_layer(gf, cur, state_copy, state_mask, il); + cur = build_mamba_layer(gf, cur, state_copy, state_mask, ubatch, il); + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // residual + cur = ggml_add(ctx0, cur, inpL); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + // final rmsnorm + cur = build_norm(inpL, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } + + // TODO: split + ggml_tensor * build_mamba_layer( + ggml_cgraph * gf, + ggml_tensor * cur, + ggml_tensor * state_copy, + ggml_tensor * state_mask, + const llama_ubatch & ubatch, + int il) const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + const auto kv_head = kv_self->head; + + const int64_t d_conv = hparams.ssm_d_conv; + const int64_t d_inner = hparams.ssm_d_inner; + const int64_t d_state = hparams.ssm_d_state; + const int64_t dt_rank = hparams.ssm_dt_rank; + const int64_t n_seqs = ubatch.n_seqs; + // Some variants of Mamba arch (e.g. FalconMamba do apply layer norm on B and Dt layers) + const bool ssm_dt_b_c_rms = hparams.ssm_dt_b_c_rms; + // Use the same RMS norm as the final layer norm + const float norm_rms_eps = hparams.f_norm_rms_eps; + + const int64_t n_seq_tokens = ubatch.n_seq_tokens; + + GGML_ASSERT(n_seqs != 0); + GGML_ASSERT(ubatch.equal_seqs); + GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); + + ggml_tensor * conv_states_all = kv_self->k_l[il]; + ggml_tensor * ssm_states_all = kv_self->v_l[il]; + + // (ab)using the KV cache to store the states + ggml_tensor * conv = build_copy_mask_state( + gf, conv_states_all, state_copy, state_mask, + hparams.n_embd_k_s(), n_seqs); + conv = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner, n_seqs); + ggml_tensor * ssm = build_copy_mask_state( + gf, ssm_states_all, state_copy, state_mask, + hparams.n_embd_v_s(), n_seqs); + ssm = ggml_reshape_3d(ctx0, ssm, d_state, d_inner, n_seqs); + + // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs} + cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs); + + // {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs} + ggml_tensor * xz = build_lora_mm(model.layers[il].ssm_in, cur); + // split the above in two + // => {d_inner, n_seq_tokens, n_seqs} + ggml_tensor * x = ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], 0); + ggml_tensor * z = ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], d_inner*ggml_element_size(xz)); + + // conv + { + // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs} + ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, x), 0); + + // copy last (d_conv - 1) columns back into the state cache + ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner, n_seqs, conv_x->nb[1], conv_x->nb[2], n_seq_tokens*(conv_x->nb[0])); + + ggml_build_forward_expand(gf, + ggml_cpy(ctx0, last_conv, + ggml_view_1d(ctx0, conv_states_all, + (d_conv - 1)*(d_inner)*(n_seqs), + kv_head*(d_conv - 1)*(d_inner)*ggml_element_size(conv_states_all)))); + + // 1D convolution + // The equivalent is to make a self-overlapping view of conv_x + // over d_conv columns at each stride in the 3rd dimension, + // then element-wise multiply that with the conv1d weight, + // then sum the elements of each row, + // (the last two steps are a dot product over rows (also doable with mul_mat)) + // then permute away the ne[0] dimension, + // and then you're left with the resulting x tensor. + // For simultaneous sequences, all sequences need to have the same length. + x = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d); + + // bias + x = ggml_add(ctx0, x, model.layers[il].ssm_conv1d_b); + + x = ggml_silu(ctx0, x); + } + + // ssm + { + // {d_inner, dt_rank + 2*d_state} @ {d_inner, n_seq_tokens, n_seqs} => {dt_rank + 2*d_state, n_seq_tokens, n_seqs} + ggml_tensor * x_db = build_lora_mm(model.layers[il].ssm_x, x); + // split + ggml_tensor * dt = ggml_view_3d(ctx0, x_db, dt_rank, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], 0); + ggml_tensor * B = ggml_view_3d(ctx0, x_db, d_state, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*dt_rank); + ggml_tensor * C = ggml_view_3d(ctx0, x_db, d_state, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*(dt_rank+d_state)); + + // Some Mamba variants (e.g. FalconMamba) apply RMS norm in B, C & Dt layers + if (ssm_dt_b_c_rms) { + dt = ggml_rms_norm(ctx0, dt, norm_rms_eps); + B = ggml_rms_norm(ctx0, B, norm_rms_eps); + C = ggml_rms_norm(ctx0, C, norm_rms_eps); + } + + // {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs} + dt = build_lora_mm(model.layers[il].ssm_dt, dt); + dt = ggml_add(ctx0, dt, model.layers[il].ssm_dt_b); + + // Custom operator to optimize the parallel associative scan + // as described in the Annex D of the Mamba paper. + // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs} + ggml_tensor * y_ssm = ggml_ssm_scan(ctx0, ssm, x, dt, model.layers[il].ssm_a, B, C); + + // store last states + ggml_build_forward_expand(gf, + ggml_cpy(ctx0, + ggml_view_1d(ctx0, y_ssm, d_state*d_inner*n_seqs, x->nb[3]), + ggml_view_1d(ctx0, ssm_states_all, d_state*d_inner*n_seqs, kv_head*d_state*d_inner*ggml_element_size(ssm_states_all)))); + + ggml_tensor * y = ggml_view_3d(ctx0, y_ssm, d_inner, n_seq_tokens, n_seqs, x->nb[1], x->nb[2], 0); + + // TODO: skip computing output earlier for unused tokens + + // {d_inner, n_seq_tokens, n_seqs} * {d_inner} => {d_inner, n_seq_tokens, n_seqs} + y = ggml_add(ctx0, y, ggml_mul(ctx0, x, model.layers[il].ssm_d)); + y = ggml_mul(ctx0, y, ggml_silu(ctx0, ggml_cont(ctx0, z))); + + // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs} + cur = build_lora_mm(model.layers[il].ssm_out, y); + } + + // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens} + cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs); + //cb(cur, "mamba_out", il); + + return cur; + } +}; + +struct llm_build_command_r : public llm_graph_context { + llm_build_command_r(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + const float f_logit_scale = hparams.f_logit_scale; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM, il); + cb(cur, "attn_norm", il); + ggml_tensor * ffn_inp = cur; + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + if (model.layers[il].attn_q_norm) { + Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens, + ggml_element_size(Qcur) * n_embd_head, + ggml_element_size(Qcur) * n_embd_head * n_head, + 0); + cb(Qcur, "Qcur", il); + Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens, + ggml_element_size(Kcur) * n_embd_head, + ggml_element_size(Kcur) * n_embd_head * n_head_kv, + 0); + cb(Kcur, "Kcur", il); + + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, + NULL, + LLM_NORM, il); + cb(Qcur, "Qcur", il); + + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, + NULL, + LLM_NORM, il); + cb(Kcur, "Kcur", il); + } + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); + } + + ggml_tensor * attn_out = cur; + + // feed-forward network + { + cur = build_ffn(ffn_inp, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + // add together residual + FFN + self-attention + cur = ggml_add(ctx0, cur, inpL); + cur = ggml_add(ctx0, cur, attn_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + if (f_logit_scale) { + cur = ggml_scale(ctx0, cur, f_logit_scale); + } + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_cohere2 : public llm_graph_context { + llm_build_cohere2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + const float f_logit_scale = hparams.f_logit_scale; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, true); + + for (int il = 0; il < n_layer; ++il) { + const bool is_sliding = hparams.is_sliding(il); + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il); + cb(cur, "attn_norm", il); + ggml_tensor * ffn_inp = cur; + + // self-attention + { + // rope freq factors for 128k context + ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + if (is_sliding) { + Qcur = ggml_rope_ext(ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, + beta_fast, beta_slow); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext(ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor, + attn_factor, beta_fast, beta_slow); + cb(Kcur, "Kcur", il); + } else { + // For non-sliding layers, just reshape without applying RoPE + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + cb(Qcur, "Qcur", il); + + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + cb(Kcur, "Kcur", il); + } + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); + } + + ggml_tensor * attn_out = cur; + + // feed-forward network + { + cur = build_ffn(ffn_inp, model.layers[il].ffn_up, NULL, NULL, model.layers[il].ffn_gate, + NULL, NULL, model.layers[il].ffn_down, NULL, NULL, NULL, LLM_FFN_SILU, LLM_FFN_PAR, + il); + cb(cur, "ffn_out", il); + } + + // add together residual + FFN + self-attention + cur = ggml_add(ctx0, cur, inpL); + cur = ggml_add(ctx0, cur, attn_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + if (f_logit_scale) { + cur = ggml_scale(ctx0, cur, f_logit_scale); + } + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +// ref: https://allenai.org/olmo +// based on the original build_llama() function, changes: +// * non-parametric layer norm +// * clamp qkv +// * removed bias +// * removed MoE +struct llm_build_olmo : public llm_graph_context { + llm_build_olmo(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + NULL, NULL, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (hparams.f_clamp_kqv > 0.0f) { + Qcur = ggml_clamp(ctx0, Qcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (hparams.f_clamp_kqv > 0.0f) { + Kcur = ggml_clamp(ctx0, Kcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (hparams.f_clamp_kqv > 0.0f) { + Vcur = ggml_clamp(ctx0, Vcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, nullptr, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + NULL, NULL, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + NULL, NULL, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_olmo2 : public llm_graph_context { + llm_build_olmo2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = inpL; + + // self_attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, + LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, + LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur_rope", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur_rope", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + cur = build_norm(cur, + model.layers[il].attn_post_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_ffn(ffn_inp, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = build_norm(cur, + model.layers[il].ffn_post_norm, NULL, + LLM_NORM_RMS, -1); + cb(cur, "ffn_post_norm", -1); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +// based on the build_qwen2moe() function, changes: +// * removed shared experts +// * removed bias +// * added q, k norm +struct llm_build_olmoe : public llm_graph_context { + llm_build_olmoe(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, + LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, + LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur_rope", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur_rope", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_openelm : public llm_graph_context { + llm_build_openelm(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + const int64_t n_head = hparams.n_head(il); + const int64_t n_head_kv = hparams.n_head_kv(il); + const int64_t n_head_qkv = 2*n_head_kv + n_head; + + cur = inpL; + ggml_tensor * residual = cur; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_reshape_3d(ctx0, cur, n_embd_head_k, n_head_qkv, n_tokens); + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, cur->nb[1], cur->nb[2], 0)); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*n_head)); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*(n_head+n_head_kv))); + cb(Vcur, "Vcur", il); + + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, NULL, + LLM_NORM_RMS, il); + cb(Qcur, "Qcur", il); + + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, NULL, + LLM_NORM_RMS, il); + cb(Kcur, "Kcur", il); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, NULL, n_rot, rope_type, n_ctx_orig, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, NULL, n_rot, rope_type, n_ctx_orig, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + Vcur = ggml_reshape_2d(ctx0, Vcur, n_embd_head * n_head_kv, n_tokens); + cb(Qcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + residual = ggml_get_rows(ctx0, residual, inp_out_ids); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + inpL = cur; + } + + cur = inpL; + + // norm + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_gptneox : public llm_graph_context { + llm_build_gptneox(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // ffn + if (hparams.use_par_res) { + // attention and ffn are computed in parallel + // x = x + attn(ln1(x)) + ffn(ln2(x)) + + ggml_tensor * attn_out = cur; + + cur = build_norm(inpL, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, inpL); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, attn_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } else { + // attention and ffn are computed sequentially + // x = x + attn(ln1(x)) + // x = x + ffn(ln2(x)) + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_arctic : public llm_graph_context { + llm_build_arctic(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp); + cb(ffn_out, "ffn_out", il); + + // MoE + cur = build_norm(inpSA, + model.layers[il].ffn_norm_exps, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm_exps", il); + + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + + cur = ggml_add(ctx0, cur, ffn_out); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_deepseek : public llm_graph_context { + llm_build_deepseek(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, kq_scale, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + if ((uint32_t) il < hparams.n_layer_dense_lead) { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + // MoE branch + ggml_tensor * moe_out = + build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + false, hparams.expert_weights_scale, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(moe_out, "ffn_moe_out", il); + + // FFN shared expert + { + ggml_tensor * ffn_shexp = build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_deepseek2 : public llm_graph_context { + llm_build_deepseek2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + bool is_lite = (hparams.n_layer == 27); + + // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly. + // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation. + const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale)); + const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(hparams.n_embd_head_k)); + const float attn_factor_scaled = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale)); + + const uint32_t n_embd_head_qk_rope = hparams.n_rot; + const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; + const uint32_t kv_lora_rank = hparams.n_lora_kv; + + ggml_tensor * cur; + ggml_tensor * inpL; + + // {n_embd, n_tokens} + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + ggml_tensor * q = NULL; + if (!is_lite) { + // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens} + q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur); + cb(q, "q", il); + + q = build_norm(q, + model.layers[il].attn_q_a_norm, NULL, + LLM_NORM_RMS, il); + cb(q, "q", il); + + // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens} + q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q); + cb(q, "q", il); + } else { + q = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(q, "q", il); + } + + // split into {n_head * n_embd_head_qk_nope, n_tokens} + ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + 0); + cb(q_nope, "q_nope", il); + + // and {n_head * n_embd_head_qk_rope, n_tokens} + ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + ggml_row_size(q->type, n_embd_head_qk_nope)); + cb(q_pe, "q_pe", il); + + // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens} + ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); + cb(kv_pe_compresseed, "kv_pe_compresseed", il); + + // split into {kv_lora_rank, n_tokens} + ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens, + kv_pe_compresseed->nb[1], + 0); + cb(kv_compressed, "kv_compressed", il); + + // and {n_embd_head_qk_rope, n_tokens} + ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens, + kv_pe_compresseed->nb[1], + kv_pe_compresseed->nb[1], + ggml_row_size(kv_pe_compresseed->type, kv_lora_rank)); + cb(k_pe, "k_pe", il); + + // TODO: the CUDA backend used to not support non-cont. (RMS) norm, investigate removing ggml_cont + kv_compressed = ggml_cont(ctx0, kv_compressed); + kv_compressed = build_norm(kv_compressed, + model.layers[il].attn_kv_a_norm, NULL, + LLM_NORM_RMS, il); + cb(kv_compressed, "kv_compressed", il); + + // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens} + ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed); + cb(kv, "kv", il); + + // split into {n_head * n_embd_head_qk_nope, n_tokens} + ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v), + ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)), + 0); + cb(k_nope, "k_nope", il); + + // and {n_head * n_embd_head_v, n_tokens} + ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens, + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)), + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head), + ggml_row_size(kv->type, (n_embd_head_qk_nope))); + cb(v_states, "v_states", il); + + v_states = ggml_cont(ctx0, v_states); + cb(v_states, "v_states", il); + + v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens, + ggml_row_size(kv->type, hparams.n_embd_head_v * n_head), + 0); + cb(v_states, "v_states", il); + + q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this + q_pe = ggml_rope_ext( + ctx0, q_pe, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor_scaled, beta_fast, beta_slow + ); + cb(q_pe, "q_pe", il); + + // shared RoPE key + k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this + k_pe = ggml_rope_ext( + ctx0, k_pe, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor_scaled, beta_fast, beta_slow + ); + cb(k_pe, "k_pe", il); + + ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0); + cb(q_states, "q_states", il); + + ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0); + cb(k_states, "k_states", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + q_states, k_states, v_states, nullptr, kq_scale, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + if ((uint32_t) il < hparams.n_layer_dense_lead) { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + // MoE branch + ggml_tensor * moe_out = + build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + model.layers[il].ffn_exp_probs_b, + n_expert, n_expert_used, + LLM_FFN_SILU, hparams.expert_weights_norm, + true, hparams.expert_weights_scale, + (llama_expert_gating_func_type) hparams.expert_gating_func, + il); + cb(moe_out, "ffn_moe_out", il); + + // FFN shared expert + { + ggml_tensor * ffn_shexp = build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_bitnet : public llm_graph_context { + llm_build_bitnet(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + if (model.layers[il].wq_scale) { + Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_scale); + } + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + // B1.K + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + if (model.layers[il].wk_scale) { + Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_scale); + } + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + // B1.V + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + if (model.layers[il].wv_scale) { + Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_scale); + } + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + NULL, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + + cur = build_norm(cur, + model.layers[il].attn_sub_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_sub_norm", il); + + cur = build_lora_mm(model.layers[il].wo, cur); + if (model.layers[il].wo_scale) { + cur = ggml_mul(ctx0, cur, model.layers[il].wo_scale); + } + if (model.layers[il].bo) { + cur = ggml_add(ctx0, cur, model.layers[il].bo); + } + cb(cur, "attn_o_out", il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward forward + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_scale, + model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_scale, + NULL, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_sub_out", il); + + cur = build_norm(cur, + model.layers[il].ffn_sub_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_sub_norm", il); + + cur = build_lora_mm(model.layers[il].ffn_down, cur); + if (model.layers[il].ffn_down_scale) { + cur = ggml_mul(ctx0, cur, model.layers[il].ffn_down_scale); + } + cb(cur, "ffn_down", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + // FIXME: do not use model.tok_embd directly, duplicate as model.output + cur = build_lora_mm(model.tok_embd, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_t5_enc : public llm_graph_context { + llm_build_t5_enc(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + ggml_tensor * pos_bucket_enc = build_inp_pos_bucket_enc(); + + auto * inp_attn = build_attn_inp_no_cache(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm_enc, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_enc, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_enc, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_enc, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc; + ggml_tensor * kq_b = build_pos_bias(pos_bucket_enc, attn_rel_b); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo_enc, nullptr, + Qcur, Kcur, Vcur, kq_b, 1.0f, il); + cb(cur, "kqv_out", il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm_enc, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // T5 uses relu, flan-T5 uses gelu-gated + cur = build_ffn(cur, + model.layers[il].ffn_up_enc, NULL, NULL, + model.layers[il].ffn_gate_enc, NULL, NULL, + model.layers[il].ffn_down_enc, NULL, NULL, + NULL, + model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU, + model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ, + il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + cb(cur, "result_embd", -1); + + cur = build_norm(cur, + model.output_norm_enc, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_t5_dec : public llm_graph_context { + llm_build_t5_dec(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + //const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + ggml_tensor * embd_enc = build_inp_cross_embd(); + ggml_tensor * pos_bucket_dec = build_inp_pos_bucket_dec(); + + const int64_t n_outputs_enc = embd_enc->ne[1]; + + auto * inp_attn_self = build_attn_inp_kv_unified(true, false); + auto * inp_attn_cross = build_attn_inp_cross(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b; + ggml_tensor * kq_b = build_pos_bias(pos_bucket_dec, attn_rel_b); + + cur = build_attn(inp_attn_self, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, kq_b, 1.0f, il); + cb(cur, "kqv_out", il); + } + + cur = ggml_add(ctx0, cur, inpSA); + cb(cur, "cross_inp", il); + + ggml_tensor * inpCA = cur; + + // norm + cur = build_norm(cur, + model.layers[il].attn_norm_cross, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm_cross", il); + + // cross-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_cross, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_cross, embd_enc); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_cross, embd_enc); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_outputs_enc); + + cur = build_attn(inp_attn_cross, gf, + model.layers[il].wo_cross, nullptr, + Qcur, Kcur, Vcur, nullptr, 1.0f, il); + cb(cur, "kqv_out", il); + + //ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); + //ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3)); + + //ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + //cb(kq, "kq", il); + + //kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias); + //cb(kq, "kq_soft_max_ext", il); + + //ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc))); + //cb(v, "v", il); + + //ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq); + //cb(kqv, "kqv", il); + + //ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); + //cb(kqv_merged, "kqv_merged", il); + + //cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); + //cb(cur, "kqv_merged_cont", il); + + //ggml_build_forward_expand(gf, cur); + + //cur = build_lora_mm(model.layers[il].wo_cross, cur); + //cb(cur, "kqv_out", il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // T5 uses relu, flan-T5 uses gelu-gated + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU, + model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ, + il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + cb(cur, "result_embd", -1); + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_jais : public llm_graph_context { + llm_build_jais(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*cur->nb[0]*(n_embd))); + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/float(n_embd_head), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + inpL = ggml_add(ctx0, cur, ffn_inp); + cb(inpL, "l_out", il); + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_chatglm : public llm_graph_context { + llm_build_chatglm(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, + NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = nullptr; + ggml_tensor * Kcur = nullptr; + ggml_tensor * Vcur = nullptr; + + if (model.layers[il].wqkv == nullptr) { + Qcur = build_lora_mm(model.layers[il].wq, cur); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + } + Kcur = build_lora_mm(model.layers[il].wk, cur); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + } + Vcur = build_lora_mm(model.layers[il].wv, cur); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + } + } else { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + if (model.layers[il].bqkv) { + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + } + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + //printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor); + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur_rope", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur_rope", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // Add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SWIGLU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + } + + inpL = ggml_add(ctx0, cur, ffn_inp); + cb(inpL, "l_out", il); + } + + cur = build_norm(inpL, + model.output_norm, + NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_nemotron : public llm_graph_context { + llm_build_nemotron(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + //GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_exaone : public llm_graph_context { + llm_build_exaone(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_rwkv6_base : public llm_graph_context { + const llama_model & model; + + llm_build_rwkv6_base(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params), model(model) { + } + + ggml_tensor * build_rwkv6_channel_mix( + const llama_layer * layer, + ggml_tensor * cur, + ggml_tensor * x_prev, + llm_arch arch) const { + ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur); + switch (arch) { + case LLM_ARCH_RWKV6: + { + ggml_tensor * xk = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_k), cur); + ggml_tensor * xr = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_r), cur); + + ggml_tensor * r = ggml_sigmoid(ctx0, build_lora_mm(layer->channel_mix_receptance, xr)); + ggml_tensor * k = ggml_sqr( + ctx0, + ggml_relu( + ctx0, + build_lora_mm(layer->channel_mix_key, xk) + ) + ); + cur = ggml_mul(ctx0, r, build_lora_mm(layer->channel_mix_value, k)); + } break; + default: + GGML_ABORT("fatal error"); + } + + return cur; + } + + ggml_tensor * build_rwkv6_time_mix( + ggml_cgraph * gf, + ggml_tensor * cur, + ggml_tensor * x_prev, + ggml_tensor * state_copy, + ggml_tensor * state_mask, + const llama_ubatch & ubatch, + int il) const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + const auto n_tokens = ubatch.n_tokens; + const auto n_seqs = ubatch.n_seqs; + const auto n_embd = hparams.n_embd; + const auto head_size = hparams.wkv_head_size; + const auto n_head = n_embd / head_size; + const auto n_head_kv = hparams.n_head_kv(il); + + const auto kv_head = kv_self->head; + + const auto & layer = model.layers[il]; + + bool is_qrwkv = layer.time_mix_first == nullptr; + + ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur); + ggml_tensor * xxx = ggml_add(ctx0, ggml_mul(ctx0, sx, layer.time_mix_lerp_x), cur); + + xxx = ggml_reshape_4d( + ctx0, + ggml_tanh( + ctx0, + ggml_mul_mat(ctx0, layer.time_mix_w1, xxx) + ), + layer.time_mix_w1->ne[1] / 5, 1, 5, n_tokens + ); + + xxx = ggml_cont(ctx0, ggml_permute(ctx0, xxx, 0, 1, 3, 2)); + + xxx = ggml_mul_mat( + ctx0, + ggml_reshape_4d( + ctx0, + layer.time_mix_w2, + layer.time_mix_w2->ne[0], layer.time_mix_w2->ne[1], 1, 5 + ), + xxx + ); + + ggml_tensor *xw, *xk, *xv, *xr, *xg; + if (layer.time_mix_lerp_fused) { + // fusing these weights makes some performance improvement + sx = ggml_reshape_3d(ctx0, sx, n_embd, 1, n_tokens); + cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens); + xxx = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xxx, layer.time_mix_lerp_fused), sx), cur); + xw = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0); + xk = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float)); + xv = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float)); + xr = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float)); + xg = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float)); + } else { + // for backward compatibility + xw = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0); + xk = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float)); + xv = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float)); + xr = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float)); + xg = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float)); + + xw = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xw, layer.time_mix_lerp_w), sx), cur); + xk = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xk, layer.time_mix_lerp_k), sx), cur); + xv = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xv, layer.time_mix_lerp_v), sx), cur); + xr = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xr, layer.time_mix_lerp_r), sx), cur); + xg = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xg, layer.time_mix_lerp_g), sx), cur); + } + + ggml_tensor * r = build_lora_mm(layer.time_mix_receptance, xr); + ggml_tensor * k = build_lora_mm(layer.time_mix_key, xk); + ggml_tensor * v = build_lora_mm(layer.time_mix_value, xv); + if (layer.time_mix_receptance_b) { + r = ggml_add(ctx0, r, layer.time_mix_receptance_b); + } + if (layer.time_mix_key_b) { + k = ggml_add(ctx0, k, layer.time_mix_key_b); + } + if (layer.time_mix_value_b) { + v = ggml_add(ctx0, v, layer.time_mix_value_b); + } + + ggml_tensor * g = build_lora_mm(layer.time_mix_gate, xg); + if (is_qrwkv) { + g = ggml_sigmoid(ctx0, g); + } else { + g = ggml_silu(ctx0, g); + } + + if (n_head_kv != 0 && n_head_kv != n_head) { + GGML_ASSERT(n_head % n_head_kv == 0); + k = ggml_reshape_4d(ctx0, k, head_size, 1, n_head_kv, n_tokens); + v = ggml_reshape_4d(ctx0, v, head_size, 1, n_head_kv, n_tokens); + ggml_tensor * tmp = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, head_size, n_head / n_head_kv, n_head_kv, n_tokens); + k = ggml_repeat(ctx0, k, tmp); + v = ggml_repeat(ctx0, v, tmp); + } + + k = ggml_reshape_3d(ctx0, k, head_size, n_head, n_tokens); + v = ggml_reshape_3d(ctx0, v, head_size, n_head, n_tokens); + r = ggml_reshape_3d(ctx0, r, head_size, n_head, n_tokens); + + ggml_tensor * w = ggml_mul_mat( + ctx0, + layer.time_mix_decay_w2, + ggml_tanh( + ctx0, + ggml_mul_mat(ctx0, layer.time_mix_decay_w1, xw) + ) + ); + + w = ggml_add(ctx0, w, layer.time_mix_decay); + w = ggml_exp(ctx0, ggml_neg(ctx0, ggml_exp(ctx0, w))); + w = ggml_reshape_3d(ctx0, w, head_size, n_head, n_tokens); + + if (is_qrwkv) { + // k = k * (1 - w) + k = ggml_sub(ctx0, k, ggml_mul(ctx0, k, w)); + } + + ggml_tensor * wkv_state = build_copy_mask_state( + gf, kv_self->v_l[il], state_copy, state_mask, + hparams.n_embd_v_s(), n_seqs); + + ggml_tensor * wkv_output; + if (is_qrwkv) { + wkv_output = ggml_gated_linear_attn(ctx0, k, v, r, w, wkv_state, pow(head_size, -0.5f)); + } else { + wkv_output = ggml_rwkv_wkv6(ctx0, k, v, r, layer.time_mix_first, w, wkv_state); + } + cur = ggml_view_1d(ctx0, wkv_output, n_embd * n_tokens, 0); + wkv_state = ggml_view_1d(ctx0, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float)); + + ggml_build_forward_expand( + gf, + ggml_cpy( + ctx0, + wkv_state, + ggml_view_1d( + ctx0, + kv_self->v_l[il], + hparams.n_embd_v_s() * n_seqs, + hparams.n_embd_v_s() * kv_head * ggml_element_size(kv_self->v_l[il]) + ) + ) + ); + + if (!is_qrwkv) { + // group norm with head_count groups + cur = ggml_reshape_3d(ctx0, cur, n_embd / n_head, n_head, n_tokens); + cur = ggml_norm(ctx0, cur, 64e-5f); + + // Convert back to regular vectors. + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.time_mix_ln), layer.time_mix_ln_b); + } else { + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + } + + cur = ggml_mul(ctx0, cur, g); + cur = build_lora_mm(layer.time_mix_output, cur); + + return cur; + } +}; + +struct llm_build_rwkv6 : public llm_build_rwkv6_base { + llm_build_rwkv6(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_build_rwkv6_base(model, params) { + GGML_ASSERT(hparams.token_shift_count == 2); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1); + + ggml_tensor * state_copy = build_inp_s_copy(); + ggml_tensor * state_mask = build_inp_s_mask(); + + const auto n_embd = hparams.n_embd; + const auto n_seq_tokens = ubatch.n_seq_tokens; + const auto n_seqs = ubatch.n_seqs; + + for (int il = 0; il < n_layer; ++il) { + const llama_layer * layer = &model.layers[il]; + + ggml_tensor * token_shift = build_rwkv_token_shift_load( + gf, state_copy, state_mask, ubatch, il + ); + + ggml_tensor * att_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0); + ggml_tensor * ffn_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], n_embd * ggml_element_size(token_shift)); + + ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM, il); + cb(att_norm, "attn_norm", il); + + ggml_tensor * x_prev = ggml_concat( + ctx0, + att_shift, + ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), + 1 + ); + + cur = build_rwkv6_time_mix(gf, att_norm, x_prev, state_copy, state_mask, ubatch, il); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + ggml_tensor * ffn_norm = build_norm(ffn_inp, layer->attn_norm_2, layer->attn_norm_2_b, LLM_NORM, il); + cb(ffn_norm, "ffn_norm", il); + + x_prev = ggml_concat( + ctx0, + ffn_shift, + ggml_view_3d(ctx0, ffn_norm, n_embd, n_seq_tokens - 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], 0), + 1 + ); + + cur = build_rwkv6_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV6); + cur = ggml_add(ctx0, cur, ffn_inp); + + token_shift = ggml_concat(ctx0, + ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm)), + ggml_view_3d(ctx0, ffn_norm, n_embd, 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(ffn_norm)), + 1 + ); + ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); + + if (hparams.rescale_every_n_layers != 0 && (il + 1) % hparams.rescale_every_n_layers == 0) { + cur = ggml_scale(ctx0, cur, 0.5F); + } + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + + cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +// ref: https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1/blob/main/modeling_rwkv6qwen2.py +struct llm_build_rwkv6qwen2 : public llm_build_rwkv6_base { + llm_build_rwkv6qwen2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_build_rwkv6_base(model, params) { + GGML_ASSERT(n_embd == hparams.n_embd_k_s()); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + ggml_tensor * state_copy = build_inp_s_copy(); + ggml_tensor * state_mask = build_inp_s_mask(); + + const auto n_embd = hparams.n_embd; + const auto n_seq_tokens = ubatch.n_seq_tokens; + const auto n_seqs = ubatch.n_seqs; + + inpL = build_inp_embd(model.tok_embd); + + for (int il = 0; il < n_layer; ++il) { + const llama_layer * layer = &model.layers[il]; + + ggml_tensor * token_shift = build_rwkv_token_shift_load( + gf, state_copy, state_mask, ubatch, il + ); + + ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il); + cb(att_norm, "attn_norm", il); + + ggml_tensor * x_prev = ggml_concat( + ctx0, + token_shift, + ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), + 1 + ); + + cur = build_rwkv6_time_mix(gf, att_norm, x_prev, state_copy, state_mask, ubatch, il); + + token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm)); + ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + + cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +// ref: https://github.com/facebookresearch/chameleon +// based on the original build_llama() function, changes: +// * qk-norm +// * swin-norm +// * removed bias +// * removed MoE +struct llm_build_chameleon : public llm_graph_context { + llm_build_chameleon(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(true, false); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + if (hparams.swin_norm) { + cur = inpL; + } else { + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + } + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + if (model.layers[il].attn_q_norm) { + Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens, + ggml_element_size(Qcur) * n_embd_head, + ggml_element_size(Qcur) * n_embd_head * n_head, + 0); + cb(Qcur, "Qcur", il); + + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, + model.layers[il].attn_q_norm_b, + LLM_NORM, il); + cb(Qcur, "Qcur", il); + } + + if (model.layers[il].attn_k_norm) { + Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens, + ggml_element_size(Kcur) * n_embd_head, + ggml_element_size(Kcur) * n_embd_head * n_head_kv, + 0); + cb(Kcur, "Kcur", il); + + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, + model.layers[il].attn_k_norm_b, + LLM_NORM, il); + cb(Kcur, "Kcur", il); + } + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, nullptr, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + + if (hparams.swin_norm) { + cur = build_norm(cur, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + } + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + if (!hparams.swin_norm) { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + } + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + if (hparams.swin_norm) { + cur = build_norm(cur, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + cb(cur, "result_output_with_img_logits", -1); + + // TODO: this suppresses the output of image tokens, which is required to enable text-only outputs. + // Needs to be removed once image outputs are supported. + int img_token_end_idx = 8196; + int img_token_start_idx = 4; + int num_img_tokens = img_token_end_idx - img_token_start_idx; + // creates 1d tensor of size num_img_tokens and values -FLT_MAX, + // which ensures that text token values are always at least larger than image token values + ggml_tensor * img_logits = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, num_img_tokens); + img_logits = ggml_clamp(ctx0, img_logits, -FLT_MAX, -FLT_MAX); + cb(img_logits, "img_logits", -1); + + cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_wavtokenizer_dec : public llm_graph_context { + llm_build_wavtokenizer_dec(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, inpL)); + + cur = ggml_conv_1d_ph(ctx0, model.conv1d, cur, 1, 1); + cur = ggml_add(ctx0, cur, model.conv1d_b); + + // posnet + for (uint32_t il = 0; il < hparams.posnet.n_layer; ++il) { + const auto & layer = model.layers[il].posnet; + + inpL = cur; + + switch (il) { + case 0: + case 1: + case 3: + case 4: + { + cur = build_norm(cur, + layer.norm1, + layer.norm1_b, + LLM_NORM_GROUP, 0); + + cur = ggml_mul(ctx0, ggml_sigmoid(ctx0, cur), cur); + + cur = ggml_conv_1d_ph(ctx0, layer.conv1, cur, 1, 1); + cur = ggml_add(ctx0, cur, layer.conv1_b); + + cur = build_norm(cur, + layer.norm2, + layer.norm2_b, + LLM_NORM_GROUP, 0); + + cur = ggml_mul(ctx0, ggml_sigmoid(ctx0, cur), cur); + + cur = ggml_conv_1d_ph(ctx0, layer.conv2, cur, 1, 1); + cur = ggml_add(ctx0, cur, layer.conv2_b); + + cur = ggml_add(ctx0, cur, inpL); + } break; + case 2: + { + cur = build_norm(cur, + layer.attn_norm, + layer.attn_norm_b, + LLM_NORM_GROUP, 0); + + ggml_tensor * q; + ggml_tensor * k; + ggml_tensor * v; + + q = ggml_conv_1d_ph(ctx0, layer.attn_q, cur, 1, 1); + k = ggml_conv_1d_ph(ctx0, layer.attn_k, cur, 1, 1); + v = ggml_conv_1d_ph(ctx0, layer.attn_v, cur, 1, 1); + + q = ggml_add(ctx0, q, layer.attn_q_b); + k = ggml_add(ctx0, k, layer.attn_k_b); + v = ggml_add(ctx0, v, layer.attn_v_b); + + q = ggml_cont(ctx0, ggml_transpose(ctx0, q)); + k = ggml_cont(ctx0, ggml_transpose(ctx0, k)); + + ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + + kq = ggml_soft_max_ext(ctx0, kq, nullptr, 1.0f/sqrtf(float(hparams.posnet.n_embd)), 0.0f); + + cur = ggml_mul_mat(ctx0, kq, v); + + cur = ggml_conv_1d_ph(ctx0, layer.attn_o, cur, 1, 1); + cur = ggml_add(ctx0, cur, layer.attn_o_b); + + cur = ggml_add(ctx0, cur, inpL); + } break; + case 5: + { + cur = build_norm(cur, + layer.norm, + layer.norm_b, + LLM_NORM_GROUP, 0); + } break; + default: GGML_ABORT("unknown posnet layer"); + }; + } + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); + + cur = build_norm(cur, + model.tok_norm, + model.tok_norm_b, + LLM_NORM, -1); + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); + + inpL = cur; + + // convnext + for (uint32_t il = 0; il < hparams.convnext.n_layer; ++il) { + const auto & layer = model.layers[il].convnext; + + cur = inpL; + + cur = ggml_conv_1d_dw_ph(ctx0, layer.dw, cur, 1, 1); + cur = ggml_add(ctx0, cur, layer.dw_b); + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); + + cur = build_norm(cur, + layer.norm, + layer.norm_b, + LLM_NORM, -1); + + cur = build_ffn(cur, + layer.pw1, layer.pw1_b, NULL, + NULL, NULL, NULL, + layer.pw2, layer.pw2_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + + cur = ggml_mul(ctx0, cur, layer.gamma); + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); + + inpL = ggml_add(ctx0, cur, inpL); + } + + cur = inpL; + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); + + cur = build_norm(cur, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + // lm_head + cur = build_lora_mm(model.output, cur); + + cur = ggml_add(ctx0, cur, model.output_b); + + cb(cur, "result_embd", -1); + res->t_embd = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +llama_memory_i * llama_model::create_memory() const { + llama_memory_i * res; + + switch (arch) { + case LLM_ARCH_RWKV6: + case LLM_ARCH_RWKV6QWEN2: + case LLM_ARCH_MAMBA: + { + res = new llama_kv_cache_unified(hparams, { + /*.get_rope_factors =*/ nullptr + }); + } break; + default: + { + res = new llama_kv_cache_unified(hparams, { + /*.get_rope_factors =*/ [this](uint32_t n_ctx_per_seq, int il) { + // choose long/short freq factors based on the context size + if (layers[il].rope_freqs != nullptr) { + return layers[il].rope_freqs; + } + + if (n_ctx_per_seq > hparams.n_ctx_orig_yarn) { + return layers[il].rope_long; + } + + return layers[il].rope_short; + } + }); + } + } + + return res; +} + +llm_graph_result_ptr llama_model::build_graph( + const llm_graph_params & params, + ggml_cgraph * gf, + llm_graph_type type) const { + std::unique_ptr llm; + + switch (arch) { + case LLM_ARCH_LLAMA: + case LLM_ARCH_MINICPM: + case LLM_ARCH_GRANITE: + case LLM_ARCH_GRANITE_MOE: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_DECI: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_BAICHUAN: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_FALCON: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_GROK: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_STARCODER: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_REFACT: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_BERT: + case LLM_ARCH_JINA_BERT_V2: + case LLM_ARCH_NOMIC_BERT: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_BLOOM: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_MPT: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_STABLELM: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_QWEN: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_QWEN2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_QWEN2VL: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_QWEN2MOE: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_PHI2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_PHI3: + case LLM_ARCH_PHIMOE: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_PLAMO: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_GPT2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_CODESHELL: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_ORION: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_INTERNLM2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_MINICPM3: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_GEMMA: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_GEMMA2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_GEMMA3: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_STARCODER2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_MAMBA: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_XVERSE: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_COMMAND_R: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_COHERE2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_DBRX: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_OLMO: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_OLMO2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_OLMOE: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_OPENELM: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_GPTNEOX: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_ARCTIC: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_DEEPSEEK: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_DEEPSEEK2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_CHATGLM: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_BITNET: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_T5: + { + switch (type) { + case LLM_GRAPH_TYPE_ENCODER: + llm = std::make_unique(*this, params, gf); + break; + case LLM_GRAPH_TYPE_DEFAULT: + case LLM_GRAPH_TYPE_DECODER: + llm = std::make_unique(*this, params, gf); + break; + default: + GGML_ABORT("invalid graph type"); + }; + } break; + //case LLM_ARCH_T5ENCODER: + // { + // llm.build_t5_enc(gf); + // } break; + case LLM_ARCH_JAIS: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_NEMOTRON: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_EXAONE: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_RWKV6: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_RWKV6QWEN2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_CHAMELEON: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_WAVTOKENIZER_DEC: + { + llm = std::make_unique(*this, params, gf); + } break; + default: + GGML_ABORT("fatal error"); + } + + // add on pooling layer + llm->build_pooling(gf, cls, cls_b, cls_out, cls_out_b); + + return std::move(llm->res); +} + // // interface implementation // -struct llama_model_params llama_model_default_params() { - struct llama_model_params result = { +llama_model_params llama_model_default_params() { + llama_model_params result = { /*.devices =*/ nullptr, /*.n_gpu_layers =*/ 0, /*.split_mode =*/ LLAMA_SPLIT_MODE_LAYER, @@ -3813,59 +11164,59 @@ struct llama_model_params llama_model_default_params() { return result; } -const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model) { +const llama_vocab * llama_model_get_vocab(const llama_model * model) { return &model->vocab; } -void llama_free_model(struct llama_model * model) { +void llama_free_model(llama_model * model) { llama_model_free(model); } -void llama_model_free(struct llama_model * model) { +void llama_model_free(llama_model * model) { delete model; } -int32_t llama_model_n_ctx_train(const struct llama_model * model) { +int32_t llama_model_n_ctx_train(const llama_model * model) { return model->hparams.n_ctx_train; } -int32_t llama_model_n_embd(const struct llama_model * model) { +int32_t llama_model_n_embd(const llama_model * model) { return model->hparams.n_embd; } -int32_t llama_model_n_layer(const struct llama_model * model) { +int32_t llama_model_n_layer(const llama_model * model) { return model->hparams.n_layer; } -int32_t llama_model_n_head(const struct llama_model * model) { +int32_t llama_model_n_head(const llama_model * model) { return model->hparams.n_head(); } -int32_t llama_model_n_head_kv(const struct llama_model * model) { +int32_t llama_model_n_head_kv(const llama_model * model) { return model->hparams.n_head_kv(); } // deprecated -int32_t llama_n_ctx_train(const struct llama_model * model) { +int32_t llama_n_ctx_train(const llama_model * model) { return llama_model_n_ctx_train(model); } // deprecated -int32_t llama_n_embd(const struct llama_model * model) { +int32_t llama_n_embd(const llama_model * model) { return llama_model_n_embd(model); } // deprecated -int32_t llama_n_layer(const struct llama_model * model) { +int32_t llama_n_layer(const llama_model * model) { return llama_model_n_layer(model); } // deprecated -int32_t llama_n_head(const struct llama_model * model) { +int32_t llama_n_head(const llama_model * model) { return llama_model_n_head(model); } -enum llama_rope_type llama_model_rope_type(const struct llama_model * model) { +llama_rope_type llama_model_rope_type(const llama_model * model) { switch (model->arch) { // these models do not use RoPE case LLM_ARCH_GPT2: @@ -3923,6 +11274,7 @@ enum llama_rope_type llama_model_rope_type(const struct llama_model * model) { case LLM_ARCH_PHIMOE: case LLM_ARCH_GEMMA: case LLM_ARCH_GEMMA2: + case LLM_ARCH_GEMMA3: case LLM_ARCH_STARCODER2: case LLM_ARCH_OPENELM: case LLM_ARCH_GPTNEOX: @@ -3943,11 +11295,11 @@ enum llama_rope_type llama_model_rope_type(const struct llama_model * model) { return LLAMA_ROPE_TYPE_NONE; } -float llama_model_rope_freq_scale_train(const struct llama_model * model) { +float llama_model_rope_freq_scale_train(const llama_model * model) { return model->hparams.rope_freq_scale_train; } -int32_t llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size) { +int32_t llama_model_meta_val_str(const llama_model * model, const char * key, char * buf, size_t buf_size) { const auto & it = model->gguf_kv.find(key); if (it == model->gguf_kv.end()) { if (buf_size > 0) { @@ -3958,11 +11310,11 @@ int32_t llama_model_meta_val_str(const struct llama_model * model, const char * return snprintf(buf, buf_size, "%s", it->second.c_str()); } -int32_t llama_model_meta_count(const struct llama_model * model) { +int32_t llama_model_meta_count(const llama_model * model) { return (int)model->gguf_kv.size(); } -int32_t llama_model_meta_key_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size) { +int32_t llama_model_meta_key_by_index(const llama_model * model, int i, char * buf, size_t buf_size) { if (i < 0 || i >= (int)model->gguf_kv.size()) { if (buf_size > 0) { buf[0] = '\0'; @@ -3974,7 +11326,7 @@ int32_t llama_model_meta_key_by_index(const struct llama_model * model, int i, c return snprintf(buf, buf_size, "%s", it->first.c_str()); } -int32_t llama_model_meta_val_str_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size) { +int32_t llama_model_meta_val_str_by_index(const llama_model * model, int32_t i, char * buf, size_t buf_size) { if (i < 0 || i >= (int)model->gguf_kv.size()) { if (buf_size > 0) { buf[0] = '\0'; @@ -3986,15 +11338,15 @@ int32_t llama_model_meta_val_str_by_index(const struct llama_model * model, int3 return snprintf(buf, buf_size, "%s", it->second.c_str()); } -int32_t llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size) { +int32_t llama_model_desc(const llama_model * model, char * buf, size_t buf_size) { return snprintf(buf, buf_size, "%s", model->desc().c_str()); } -uint64_t llama_model_size(const struct llama_model * model) { +uint64_t llama_model_size(const llama_model * model) { return model->size(); } -const char * llama_model_chat_template(const struct llama_model * model, const char * name) { +const char * llama_model_chat_template(const llama_model * model, const char * name) { const auto key = name ? LLM_KV(model->arch, name)(LLM_KV_TOKENIZER_CHAT_TEMPLATE_N) : LLM_KV(model->arch)(LLM_KV_TOKENIZER_CHAT_TEMPLATE); const auto & it = model->gguf_kv.find(key); @@ -4005,11 +11357,11 @@ const char * llama_model_chat_template(const struct llama_model * model, const c return it->second.c_str(); } -uint64_t llama_model_n_params(const struct llama_model * model) { +uint64_t llama_model_n_params(const llama_model * model) { return model->n_elements(); } -bool llama_model_has_encoder(const struct llama_model * model) { +bool llama_model_has_encoder(const llama_model * model) { switch (model->arch) { case LLM_ARCH_T5: return true; case LLM_ARCH_T5ENCODER: return true; @@ -4017,22 +11369,26 @@ bool llama_model_has_encoder(const struct llama_model * model) { } } -bool llama_model_has_decoder(const struct llama_model * model) { +bool llama_model_has_decoder(const llama_model * model) { switch (model->arch) { case LLM_ARCH_T5ENCODER: return false; default: return true; } } -llama_token llama_model_decoder_start_token(const struct llama_model * model) { +llama_token llama_model_decoder_start_token(const llama_model * model) { return model->hparams.dec_start_token_id; } -bool llama_model_is_recurrent(const struct llama_model * model) { +bool llama_model_is_recurrent(const llama_model * model) { switch (model->arch) { - case LLM_ARCH_MAMBA: return true; - case LLM_ARCH_RWKV6: return true; - case LLM_ARCH_RWKV6QWEN2: return true; - default: return false; + case LLM_ARCH_MAMBA: return true; + case LLM_ARCH_RWKV6: return true; + case LLM_ARCH_RWKV6QWEN2: return true; + default: return false; } } + +const std::vector> & llama_internal_get_tensor_map(const llama_model * model) { + return model->tensors_by_name; +} diff --git a/src/llama-model.h b/src/llama-model.h index a7c3044478..55c26a92b0 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -2,7 +2,9 @@ #include "llama.h" #include "llama-arch.h" +#include "llama-graph.h" #include "llama-hparams.h" +#include "llama-memory.h" #include "llama-vocab.h" #include @@ -10,6 +12,8 @@ #include #include +struct llama_cparams; +struct llama_ubatch; struct llama_model_loader; // available models @@ -347,7 +351,7 @@ struct llama_model { std::string desc() const; size_t size() const; - size_t max_nodes() const; + size_t n_tensors() const; size_t n_devices() const; // total number of parameters in the model @@ -362,9 +366,22 @@ struct llama_model { const struct ggml_tensor * get_tensor(const char * name) const; + // TODO: move this to new llm_arch_model_i interface + llama_memory_i * create_memory() const; // TODO: params + + // TODO: move this to new llm_arch_model_i interface + llm_graph_result_ptr build_graph( + const llm_graph_params & params, + ggml_cgraph * gf, + llm_graph_type type) const; + private: struct impl; std::unique_ptr pimpl; }; const char * llm_type_name(llm_type type); + +// For internal test use +// TODO: remove +const std::vector> & llama_internal_get_tensor_map(const llama_model * model); diff --git a/src/llama.cpp b/src/llama.cpp index 607f278615..81e1dd1d08 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -2,9370 +2,28 @@ #include "llama-chat.h" #include "llama-mmap.h" -#include "llama-context.h" #include "llama-vocab.h" -#include "llama-sampling.h" -#include "llama-kv-cache.h" #include "llama-model-loader.h" #include "llama-model.h" #include "ggml.h" -#include "ggml-alloc.h" #include "ggml-backend.h" -#include "ggml-cpp.h" #include -#include -#include -#include -#include #include #include #include #include #include -#include #if defined(_MSC_VER) #pragma warning(disable: 4244 4267) // possible loss of data #endif -// Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback -static int llama_model_load(const std::string & fname, std::vector & splits, llama_model & model, llama_model_params & params) { - // loading time will be recalculated after the first eval, so - // we take page faults deferred by mmap() into consideration - model.t_load_us = 0; - time_meas tm(model.t_load_us); - - model.t_start_us = tm.t_start_us; - - try { - llama_model_loader ml(fname, splits, params.use_mmap, params.check_tensors, params.kv_overrides); - - ml.print_info(); - - model.hparams.vocab_only = params.vocab_only; - - try { - model.load_arch(ml); - } catch(const std::exception & e) { - throw std::runtime_error("error loading model architecture: " + std::string(e.what())); - } - try { - model.load_hparams(ml); - } catch(const std::exception & e) { - throw std::runtime_error("error loading model hyperparameters: " + std::string(e.what())); - } - try { - model.load_vocab(ml); - } catch(const std::exception & e) { - throw std::runtime_error("error loading model vocabulary: " + std::string(e.what())); - } - - model.load_stats(ml); - model.print_info(); - - if (params.vocab_only) { - LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__); - return 0; - } - - if (!model.load_tensors(ml)) { - return -2; - } - } catch (const std::exception & err) { - LLAMA_LOG_ERROR("%s: error loading model: %s\n", __func__, err.what()); - return -1; - } - - return 0; -} - -// -// llm_build -// - -using llm_build_cb = std::function; - -enum llm_ffn_op_type { - LLM_FFN_SILU, - LLM_FFN_GELU, - LLM_FFN_RELU, - LLM_FFN_RELU_SQR, - LLM_FFN_SWIGLU, -}; - -enum llm_ffn_gate_type { - LLM_FFN_SEQ, - LLM_FFN_PAR, // ffn_gate is parallel to ffn_up -}; - -enum llm_norm_type { - LLM_NORM, - LLM_NORM_RMS, - LLM_NORM_GROUP, -}; - -static struct ggml_tensor * llm_build_inp_embd( - struct ggml_context * ctx, - struct llama_context & lctx, - const llama_hparams & hparams, - const llama_ubatch & ubatch, - struct ggml_tensor * tok_embd, - const llm_build_cb & cb) { - const int64_t n_embd = hparams.n_embd; - - struct ggml_tensor * inpL; - - if (ubatch.token) { - lctx.inp_tokens = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, ubatch.n_tokens); - cb(lctx.inp_tokens, "inp_tokens", -1); - ggml_set_input(lctx.inp_tokens); - - inpL = ggml_get_rows(ctx, tok_embd, lctx.inp_tokens); - - // apply lora for embedding tokens if needed - for (auto & it : lctx.lora) { - struct llama_adapter_lora_weight * lw = it.first->get_weight(tok_embd); - if (lw == nullptr) { - continue; - } - const float adapter_scale = it.second; - const float scale = lw->get_scale(it.first->alpha, adapter_scale); - struct ggml_tensor * inpL_delta = ggml_scale(ctx, ggml_mul_mat( - ctx, lw->b, // non-transposed lora_b - ggml_get_rows(ctx, lw->a, lctx.inp_tokens) - ), scale); - inpL = ggml_add(ctx, inpL, inpL_delta); - } - } else { - lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, ubatch.n_tokens); - inpL = lctx.inp_embd; - ggml_set_input(lctx.inp_embd); - } - - // For Granite architecture - if (hparams.f_embedding_scale != 0.0f) { - inpL = ggml_scale(ctx, inpL, hparams.f_embedding_scale); - } - - cb(inpL, "inp_embd", -1); - - return inpL; -} - -static void llm_build_kv_store( - struct ggml_context * ctx, - const llama_hparams & hparams, - const llama_cparams & cparams, - const llama_kv_cache & kv, - struct ggml_cgraph * graph, - struct ggml_tensor * k_cur, - struct ggml_tensor * v_cur, - int32_t n_tokens, - int32_t kv_head, - const llm_build_cb & cb, - int64_t il) { - const int64_t n_ctx = cparams.n_ctx; - - const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); - const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); - - GGML_ASSERT(kv.size == n_ctx); - - struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv.k_l[il], n_tokens*n_embd_k_gqa, ggml_row_size(kv.k_l[il]->type, n_embd_k_gqa)*kv_head); - cb(k_cache_view, "k_cache_view", il); - - // note: storing RoPE-ed version of K in the KV cache - ggml_build_forward_expand(graph, ggml_cpy(ctx, k_cur, k_cache_view)); - - assert(v_cur->ne[0] == n_embd_v_gqa && v_cur->ne[1] == n_tokens); - - struct ggml_tensor * v_cache_view = nullptr; - - if (cparams.flash_attn) { - v_cache_view = ggml_view_1d(ctx, kv.v_l[il], n_tokens*n_embd_v_gqa, ggml_row_size(kv.v_l[il]->type, n_embd_v_gqa)*kv_head); - } else { - // note: the V cache is transposed when not using flash attention - v_cache_view = ggml_view_2d(ctx, kv.v_l[il], n_tokens, n_embd_v_gqa, - ( n_ctx)*ggml_element_size(kv.v_l[il]), - (kv_head)*ggml_element_size(kv.v_l[il])); - - v_cur = ggml_transpose(ctx, v_cur); - } - cb(v_cache_view, "v_cache_view", il); - - ggml_build_forward_expand(graph, ggml_cpy(ctx, v_cur, v_cache_view)); -} - -// do mat_mul, while optionally apply lora -static struct ggml_tensor * llm_build_lora_mm( - struct llama_context & lctx, - struct ggml_context * ctx0, - struct ggml_tensor * w, - struct ggml_tensor * cur) { - struct ggml_tensor * res = ggml_mul_mat(ctx0, w, cur); - for (auto & it : lctx.lora) { - struct llama_adapter_lora_weight * lw = it.first->get_weight(w); - if (lw == nullptr) { - continue; - } - const float adapter_scale = it.second; - const float scale = lw->get_scale(it.first->alpha, adapter_scale); - struct ggml_tensor * ab_cur = ggml_mul_mat( - ctx0, lw->b, - ggml_mul_mat(ctx0, lw->a, cur) - ); - ab_cur = ggml_scale(ctx0, ab_cur, scale); - res = ggml_add(ctx0, res, ab_cur); - } - return res; -} - -// do mat_mul_id, while optionally apply lora -static struct ggml_tensor * llm_build_lora_mm_id( - struct llama_context & lctx, - struct ggml_context * ctx0, - struct ggml_tensor * w, // struct ggml_tensor * as - struct ggml_tensor * cur, // struct ggml_tensor * b - struct ggml_tensor * ids) { - struct ggml_tensor * res = ggml_mul_mat_id(ctx0, w, cur, ids); - for (auto & it : lctx.lora) { - struct llama_adapter_lora_weight * lw = it.first->get_weight(w); - if (lw == nullptr) { - continue; - } - const float alpha = it.first->alpha; - const float rank = (float) lw->b->ne[0]; - const float scale = alpha ? it.second * alpha / rank : it.second; - struct ggml_tensor * ab_cur = ggml_mul_mat_id( - ctx0, lw->b, - ggml_mul_mat_id(ctx0, lw->a, cur, ids), - ids - ); - ab_cur = ggml_scale(ctx0, ab_cur, scale); - res = ggml_add(ctx0, res, ab_cur); - } - return res; -} - -static struct ggml_tensor * llm_build_norm( - struct ggml_context * ctx, - struct ggml_tensor * cur, - const llama_hparams & hparams, - struct ggml_tensor * mw, - struct ggml_tensor * mb, - llm_norm_type type, - const llm_build_cb & cb, - int il) { - switch (type) { - case LLM_NORM: cur = ggml_norm (ctx, cur, hparams.f_norm_eps); break; - case LLM_NORM_RMS: cur = ggml_rms_norm (ctx, cur, hparams.f_norm_rms_eps); break; - case LLM_NORM_GROUP: - { - cur = ggml_reshape_3d(ctx, cur, cur->ne[0], 1, cur->ne[1]); - cur = ggml_group_norm(ctx, cur, hparams.n_norm_groups, hparams.f_norm_group_eps); - cur = ggml_reshape_2d(ctx, cur, cur->ne[0], cur->ne[2]); - } break; - } - - if (mw || mb) { - cb(cur, "norm", il); - } - - if (mw) { - cur = ggml_mul(ctx, cur, mw); - if (mb) { - cb(cur, "norm_w", il); - } - } - - if (mb) { - cur = ggml_add(ctx, cur, mb); - } - - return cur; -} - -static struct ggml_tensor * llm_build_ffn( - struct ggml_context * ctx, - struct llama_context & lctx, - struct ggml_tensor * cur, - struct ggml_tensor * up, - struct ggml_tensor * up_b, - struct ggml_tensor * up_s, - struct ggml_tensor * gate, - struct ggml_tensor * gate_b, - struct ggml_tensor * gate_s, - struct ggml_tensor * down, - struct ggml_tensor * down_b, - struct ggml_tensor * down_s, - struct ggml_tensor * act_scales, - llm_ffn_op_type type_op, - llm_ffn_gate_type type_gate, - const llm_build_cb & cb, - int il) { - struct ggml_tensor * tmp = up ? llm_build_lora_mm(lctx, ctx, up, cur) : cur; - cb(tmp, "ffn_up", il); - - if (up_b) { - tmp = ggml_add(ctx, tmp, up_b); - cb(tmp, "ffn_up_b", il); - } - - if (up_s) { - tmp = ggml_mul(ctx, tmp, up_s); - cb(tmp, "ffn_up_s", il); - } - - if (gate) { - switch (type_gate) { - case LLM_FFN_SEQ: - { - cur = llm_build_lora_mm(lctx, ctx, gate, tmp); - cb(cur, "ffn_gate", il); - } break; - case LLM_FFN_PAR: - { - cur = llm_build_lora_mm(lctx, ctx, gate, cur); - cb(cur, "ffn_gate", il); - } break; - } - - if (gate_b) { - cur = ggml_add(ctx, cur, gate_b); - cb(cur, "ffn_gate_b", il); - } - - if (gate_s) { - cur = ggml_mul(ctx, cur, gate_s); - cb(cur, "ffn_gate_s", il); - } - - } else { - cur = tmp; - } - - switch (type_op) { - case LLM_FFN_SILU: - { - cur = ggml_silu(ctx, cur); - cb(cur, "ffn_silu", il); - } break; - case LLM_FFN_GELU: - { - cur = ggml_gelu(ctx, cur); - cb(cur, "ffn_gelu", il); - if (act_scales != NULL) { - cur = ggml_div(ctx, cur, act_scales); - cb(cur, "ffn_act", il); - } - } break; - case LLM_FFN_RELU: - { - cur = ggml_relu(ctx, cur); - cb(cur, "ffn_relu", il); - } break; - case LLM_FFN_RELU_SQR: - { - cur = ggml_relu(ctx, cur); - cb(cur, "ffn_relu", il); - - cur = ggml_sqr(ctx, cur); - cb(cur, "ffn_sqr(relu)", il); - } break; - case LLM_FFN_SWIGLU: - { - // Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf - int64_t split_point = cur->ne[0] / 2; - struct ggml_tensor * x0 = ggml_cont(ctx, ggml_view_2d(ctx, cur, split_point, cur->ne[1], cur->nb[1], 0)); - struct ggml_tensor * x1 = ggml_cont(ctx, ggml_view_2d(ctx, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur))); - - x0 = ggml_silu(ctx, x0); - cb(cur, "ffn_silu", il); - - cur = ggml_mul(ctx, x0, x1); - cb(cur, "ffn_mul", il); - } break; - } - - if (type_gate == LLM_FFN_PAR) { - cur = ggml_mul(ctx, cur, tmp); - cb(cur, "ffn_gate_par", il); - } - - if (down) { - cur = llm_build_lora_mm(lctx, ctx, down, cur); - } - - if (down_b) { - cb(cur, "ffn_down", il); - } - - if (down_b) { - cur = ggml_add(ctx, cur, down_b); - } - - if (down_s) { - cur = ggml_mul(ctx, cur, down_s); - cb(cur, "ffn_down_s", il); - } - - return cur; -} - -static struct ggml_tensor * llm_build_moe_ffn( - struct ggml_context * ctx, - struct llama_context & lctx, - struct ggml_tensor * cur, - struct ggml_tensor * gate_inp, - struct ggml_tensor * up_exps, - struct ggml_tensor * gate_exps, - struct ggml_tensor * down_exps, - struct ggml_tensor * exp_probs_b, - int64_t n_expert, - int64_t n_expert_used, - llm_ffn_op_type type_op, - bool norm_w, - bool scale_w, - float w_scale, -llama_expert_gating_func_type gating_op, - const llm_build_cb & cb, - int il) { - int64_t n_embd = cur->ne[0]; - int64_t n_tokens = cur->ne[1]; - - ggml_tensor * logits = llm_build_lora_mm(lctx, ctx, gate_inp, cur); // [n_expert, n_tokens] - cb(logits, "ffn_moe_logits", il); - - ggml_tensor * probs = nullptr; - switch (gating_op) { - case LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX: - { - probs = ggml_soft_max(ctx, logits); // [n_expert, n_tokens] - } break; - case LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID: - { - probs = ggml_sigmoid(ctx, logits); // [n_expert, n_tokens] - } break; - default: - GGML_ABORT("fatal error"); - } - cb(probs, "ffn_moe_probs", il); - - // add experts selection bias - introduced in DeepSeek V3 - // leave probs unbiased as it's later used to get expert weights - ggml_tensor * selection_probs = probs; - if (exp_probs_b != nullptr) { - selection_probs = ggml_add(ctx, probs, exp_probs_b); - cb(selection_probs, "ffn_moe_probs_biased", il); - } - - // select experts - ggml_tensor * selected_experts = ggml_top_k(ctx, selection_probs, n_expert_used); // [n_expert_used, n_tokens] - cb(selected_experts->src[0], "ffn_moe_argsort", il); - cb(selected_experts, "ffn_moe_topk", il); - - ggml_tensor * weights = ggml_get_rows(ctx, - ggml_reshape_3d(ctx, probs, 1, n_expert, n_tokens), selected_experts); // [1, n_expert_used, n_tokens] - cb(weights, "ffn_moe_weights", il); - - if (norm_w) { - weights = ggml_reshape_2d(ctx, weights, n_expert_used, n_tokens); - - ggml_tensor * weights_sum = ggml_sum_rows(ctx, weights); // [1, n_tokens] - cb(weights_sum, "ffn_moe_weights_sum", il); - - weights = ggml_div(ctx, weights, weights_sum); // [n_expert_used, n_tokens] - cb(weights, "ffn_moe_weights_norm", il); - - weights = ggml_reshape_3d(ctx, weights, 1, n_expert_used, n_tokens); - } - if (scale_w) { - weights = ggml_scale(ctx, weights, w_scale); - cb(weights, "ffn_moe_weights_scaled", il); - } - - cur = ggml_reshape_3d(ctx, cur, n_embd, 1, n_tokens); - ggml_tensor * up = llm_build_lora_mm_id(lctx, ctx, up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens] - cb(up, "ffn_moe_up", il); - - ggml_tensor * gate = llm_build_lora_mm_id(lctx, ctx, gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens] - cb(gate, "ffn_moe_gate", il); - - switch (type_op) { - case LLM_FFN_SILU: - { - gate = ggml_silu(ctx, gate); - cb(gate, "ffn_moe_silu", il); - } break; - case LLM_FFN_GELU: - { - gate = ggml_gelu(ctx, gate); - cb(gate, "ffn_moe_gelu", il); - } break; - default: - GGML_ABORT("fatal error"); - } - - ggml_tensor * par = ggml_mul(ctx, up, gate); // [n_ff, n_expert_used, n_tokens] - cb(par, "ffn_moe_gate_par", il); - - ggml_tensor * experts = llm_build_lora_mm_id(lctx, ctx, down_exps, par, selected_experts); // [n_embd, n_expert_used, n_tokens] - cb(experts, "ffn_moe_down", il); - - experts = ggml_mul(ctx, experts, weights); - - // aggregate experts - ggml_tensor * moe_out = nullptr; - for (int i = 0; i < n_expert_used; ++i) { - ggml_tensor * cur_expert = ggml_view_2d(ctx, experts, n_embd, n_tokens, - experts->nb[2], i*experts->nb[1]); - - if (i == 0) { - moe_out = cur_expert; - } else { - moe_out = ggml_add(ctx, moe_out, cur_expert); - } - } - - if (n_expert_used == 1) { - // avoid returning a non-contiguous tensor - moe_out = ggml_cont(ctx, moe_out); - } - - return moe_out; -} - -static struct ggml_tensor * llm_build_kqv( - struct ggml_context * ctx, - struct llama_context & lctx, - const llama_kv_cache & kv, - struct ggml_cgraph * graph, - struct ggml_tensor * wo, - struct ggml_tensor * wo_b, - struct ggml_tensor * q_cur, - struct ggml_tensor * kq_mask, - int32_t n_tokens, - int32_t n_kv, - float kq_scale, - const llm_build_cb & cb, - int il) { - const llama_model & model = lctx.model; - const llama_hparams & hparams = lctx.model.hparams; - const llama_cparams & cparams = lctx.cparams; - - const int64_t n_ctx = cparams.n_ctx; - const int64_t n_head = hparams.n_head(il); - const int64_t n_head_kv = hparams.n_head_kv(il); - const int64_t n_embd_head_k = hparams.n_embd_head_k; - const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); - const int64_t n_embd_head_v = hparams.n_embd_head_v; - const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); - - struct ggml_tensor * q = ggml_permute(ctx, q_cur, 0, 2, 1, 3); - cb(q, "q", il); - - struct ggml_tensor * k = - ggml_view_3d(ctx, kv.k_l[il], - n_embd_head_k, n_kv, n_head_kv, - ggml_row_size(kv.k_l[il]->type, n_embd_k_gqa), - ggml_row_size(kv.k_l[il]->type, n_embd_head_k), - 0); - cb(k, "k", il); - - struct ggml_tensor * cur; - - if (cparams.flash_attn) { - GGML_UNUSED(model); - GGML_UNUSED(n_ctx); - - // split cached v into n_head heads (not transposed) - struct ggml_tensor * v = - ggml_view_3d(ctx, kv.v_l[il], - n_embd_head_v, n_kv, n_head_kv, - ggml_row_size(kv.v_l[il]->type, n_embd_v_gqa), - ggml_row_size(kv.v_l[il]->type, n_embd_head_v), - 0); - cb(v, "v", il); - - cur = ggml_flash_attn_ext(ctx, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias, - hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f); - - ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32); - - cur = ggml_reshape_2d(ctx, cur, n_embd_head_v*n_head, n_tokens); - } else { - struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q); - cb(kq, "kq", il); - - // note: this op tends to require high floating point range - // while for some models F16 is enough, for others it is not, so we default to F32 here - ggml_mul_mat_set_prec(kq, GGML_PREC_F32); - - if (model.arch == LLM_ARCH_GROK) { - // need to do the following: - // multiply by attn_output_multiplyer of 0.08838834764831845 - // and then : - // kq = 30 * tanh(kq / 30) - // before the softmax below - - kq = ggml_tanh(ctx, ggml_scale(ctx, kq, 0.08838834764831845f/30.0f)); - kq = ggml_scale(ctx, kq, 30); - } - - if (hparams.attn_soft_cap) { - kq = ggml_scale(ctx, kq, 1.0f / hparams.f_attn_logit_softcapping); - kq = ggml_tanh(ctx, kq); - kq = ggml_scale(ctx, kq, hparams.f_attn_logit_softcapping); - } - - kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias); - cb(kq, "kq_soft_max_ext", il); - - GGML_ASSERT(kv.size == n_ctx); - - // split cached v into n_head heads - struct ggml_tensor * v = - ggml_view_3d(ctx, kv.v_l[il], - n_kv, n_embd_head_v, n_head_kv, - ggml_element_size(kv.v_l[il])*n_ctx, - ggml_element_size(kv.v_l[il])*n_ctx*n_embd_head_v, - 0); - cb(v, "v", il); - - struct ggml_tensor * kqv = ggml_mul_mat(ctx, v, kq); - cb(kqv, "kqv", il); - - struct ggml_tensor * kqv_merged = ggml_permute(ctx, kqv, 0, 2, 1, 3); - cb(kqv_merged, "kqv_merged", il); - - cur = ggml_cont_2d(ctx, kqv_merged, n_embd_head_v*n_head, n_tokens); - cb(cur, "kqv_merged_cont", il); - } - - ggml_build_forward_expand(graph, cur); - - if (wo) { - cur = llm_build_lora_mm(lctx, ctx, wo, cur); - } - - if (wo_b) { - cb(cur, "kqv_wo", il); - } - - if (wo_b) { - cur = ggml_add(ctx, cur, wo_b); - } - - return cur; -} - -static struct ggml_tensor * llm_build_kv( - struct ggml_context * ctx, - struct llama_context & lctx, - const llama_kv_cache & kv, - struct ggml_cgraph * graph, - struct ggml_tensor * wo, - struct ggml_tensor * wo_b, - struct ggml_tensor * k_cur, - struct ggml_tensor * v_cur, - struct ggml_tensor * q_cur, - struct ggml_tensor * kq_mask, - int32_t n_tokens, - int32_t kv_head, - int32_t n_kv, - float kq_scale, - const llm_build_cb & cb, - int il) { - const llama_hparams & hparams = lctx.model.hparams; - const llama_cparams & cparams = lctx.cparams; - - // these nodes are added to the graph together so that they are not reordered - // by doing so, the number of splits in the graph is reduced - ggml_build_forward_expand(graph, q_cur); - ggml_build_forward_expand(graph, k_cur); - ggml_build_forward_expand(graph, v_cur); - - llm_build_kv_store(ctx, hparams, cparams, kv, graph, k_cur, v_cur, n_tokens, kv_head, cb, il); - - struct ggml_tensor * cur; - - cur = llm_build_kqv(ctx, lctx, kv, graph, wo, wo_b, q_cur, kq_mask, n_tokens, n_kv, kq_scale, cb, il); - cb(cur, "kqv_out", il); - - return cur; -} - -static struct ggml_tensor * llm_build_copy_mask_state( - struct ggml_context * ctx, - struct ggml_cgraph * graph, - struct ggml_tensor * s, - struct ggml_tensor * state_copy, - struct ggml_tensor * state_mask, - int32_t n_state, - int32_t kv_size, - int32_t kv_head, - int32_t n_kv, - int32_t n_seqs) { - struct ggml_tensor * states = ggml_reshape_2d(ctx, s, n_state, kv_size); - - // copy states - // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv - // this shrinks the tensors's ne[1] to n_kv - states = ggml_get_rows(ctx, states, state_copy); - - // clear states of sequences which are starting at the beginning of this batch - // FIXME: zero-out NANs? - states = ggml_mul(ctx, states, state_mask); - - // copy states which won't be changed further (between n_seqs and n_kv) - ggml_build_forward_expand(graph, - ggml_cpy(ctx, - ggml_view_1d(ctx, states, n_state*(n_kv - n_seqs), n_seqs*n_state*ggml_element_size(states)), - ggml_view_1d(ctx, s, n_state*(n_kv - n_seqs), (kv_head + n_seqs)*n_state*ggml_element_size(s)))); - - // the part of the states that will be used and modified - return ggml_view_2d(ctx, states, n_state, n_seqs, states->nb[1], 0); -} - -// TODO: split -static struct ggml_tensor * llm_build_mamba( - struct ggml_context * ctx, - struct llama_context & lctx, - const llama_ubatch & ubatch, - struct ggml_cgraph * graph, - struct ggml_tensor * cur, - struct ggml_tensor * state_copy, - struct ggml_tensor * state_mask, - int32_t kv_head, - int32_t n_kv, - const llm_build_cb & cb, - int il) { - const llama_model & model = lctx.model; - const llama_hparams & hparams = model.hparams; - const llama_kv_cache & kv = lctx.kv_self; - const int64_t d_conv = hparams.ssm_d_conv; - const int64_t d_inner = hparams.ssm_d_inner; - const int64_t d_state = hparams.ssm_d_state; - const int64_t dt_rank = hparams.ssm_dt_rank; - const int64_t n_seqs = ubatch.n_seqs; - // Some variants of Mamba arch (e.g. FalconMamba do apply layer norm on B and Dt layers) - const bool ssm_dt_b_c_rms = hparams.ssm_dt_b_c_rms; - // Use the same RMS norm as the final layer norm - const float norm_rms_eps = hparams.f_norm_rms_eps; - - const int64_t n_seq_tokens = ubatch.n_seq_tokens; - - GGML_ASSERT(n_seqs != 0); - GGML_ASSERT(ubatch.equal_seqs); - GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); - - struct ggml_tensor * conv_states_all = kv.k_l[il]; - struct ggml_tensor * ssm_states_all = kv.v_l[il]; - - // (ab)using the KV cache to store the states - struct ggml_tensor * conv = llm_build_copy_mask_state(ctx, - graph, conv_states_all, state_copy, state_mask, - hparams.n_embd_k_s(), kv.size, kv_head, n_kv, n_seqs); - conv = ggml_reshape_3d(ctx, conv, d_conv - 1, d_inner, n_seqs); - struct ggml_tensor * ssm = llm_build_copy_mask_state(ctx, - graph, ssm_states_all, state_copy, state_mask, - hparams.n_embd_v_s(), kv.size, kv_head, n_kv, n_seqs); - ssm = ggml_reshape_3d(ctx, ssm, d_state, d_inner, n_seqs); - - // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs} - cur = ggml_reshape_3d(ctx, cur, cur->ne[0], n_seq_tokens, n_seqs); - - // {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs} - struct ggml_tensor * xz = llm_build_lora_mm(lctx, ctx, model.layers[il].ssm_in, cur); - // split the above in two - // => {d_inner, n_seq_tokens, n_seqs} - struct ggml_tensor * x = ggml_view_3d(ctx, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], 0); - struct ggml_tensor * z = ggml_view_3d(ctx, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], d_inner*ggml_element_size(xz)); - - // conv - { - // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs} - struct ggml_tensor * conv_x = ggml_concat(ctx, conv, ggml_transpose(ctx, x), 0); - - // copy last (d_conv - 1) columns back into the state cache - struct ggml_tensor * last_conv = ggml_view_3d(ctx, conv_x, d_conv - 1, d_inner, n_seqs, conv_x->nb[1], conv_x->nb[2], n_seq_tokens*(conv_x->nb[0])); - - ggml_build_forward_expand(graph, - ggml_cpy(ctx, last_conv, - ggml_view_1d(ctx, conv_states_all, - (d_conv - 1)*(d_inner)*(n_seqs), - kv_head*(d_conv - 1)*(d_inner)*ggml_element_size(conv_states_all)))); - - // 1D convolution - // The equivalent is to make a self-overlapping view of conv_x - // over d_conv columns at each stride in the 3rd dimension, - // then element-wise multiply that with the conv1d weight, - // then sum the elements of each row, - // (the last two steps are a dot product over rows (also doable with mul_mat)) - // then permute away the ne[0] dimension, - // and then you're left with the resulting x tensor. - // For simultaneous sequences, all sequences need to have the same length. - x = ggml_ssm_conv(ctx, conv_x, model.layers[il].ssm_conv1d); - - // bias - x = ggml_add(ctx, x, model.layers[il].ssm_conv1d_b); - - x = ggml_silu(ctx, x); - } - - // ssm - { - // {d_inner, dt_rank + 2*d_state} @ {d_inner, n_seq_tokens, n_seqs} => {dt_rank + 2*d_state, n_seq_tokens, n_seqs} - struct ggml_tensor * x_db = llm_build_lora_mm(lctx, ctx, model.layers[il].ssm_x, x); - // split - struct ggml_tensor * dt = ggml_view_3d(ctx, x_db, dt_rank, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], 0); - struct ggml_tensor * B = ggml_view_3d(ctx, x_db, d_state, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*dt_rank); - struct ggml_tensor * C = ggml_view_3d(ctx, x_db, d_state, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*(dt_rank+d_state)); - - // Some Mamba variants (e.g. FalconMamba) apply RMS norm in B, C & Dt layers - if (ssm_dt_b_c_rms) { - dt = ggml_rms_norm(ctx, dt, norm_rms_eps); - B = ggml_rms_norm(ctx, B, norm_rms_eps); - C = ggml_rms_norm(ctx, C, norm_rms_eps); - } - - // {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs} - dt = llm_build_lora_mm(lctx, ctx, model.layers[il].ssm_dt, dt); - dt = ggml_add(ctx, dt, model.layers[il].ssm_dt_b); - - // Custom operator to optimize the parallel associative scan - // as described in the Annex D of the Mamba paper. - // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs} - struct ggml_tensor * y_ssm = ggml_ssm_scan(ctx, ssm, x, dt, model.layers[il].ssm_a, B, C); - - // store last states - ggml_build_forward_expand(graph, - ggml_cpy(ctx, - ggml_view_1d(ctx, y_ssm, d_state*d_inner*n_seqs, x->nb[3]), - ggml_view_1d(ctx, ssm_states_all, d_state*d_inner*n_seqs, kv_head*d_state*d_inner*ggml_element_size(ssm_states_all)))); - - struct ggml_tensor * y = ggml_view_3d(ctx, y_ssm, d_inner, n_seq_tokens, n_seqs, x->nb[1], x->nb[2], 0); - - // TODO: skip computing output earlier for unused tokens - - // {d_inner, n_seq_tokens, n_seqs} * {d_inner} => {d_inner, n_seq_tokens, n_seqs} - y = ggml_add(ctx, y, ggml_mul(ctx, x, model.layers[il].ssm_d)); - y = ggml_mul(ctx, y, ggml_silu(ctx, ggml_cont(ctx, z))); - - // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs} - cur = llm_build_lora_mm(lctx, ctx, model.layers[il].ssm_out, y); - } - - // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens} - cur = ggml_reshape_2d(ctx, cur, cur->ne[0], n_seq_tokens * n_seqs); - cb(cur, "mamba_out", il); - - return cur; -} - -static struct ggml_tensor * llm_build_rwkv6_time_mix( - struct llama_context & lctx, - struct ggml_context * ctx, - const struct llama_layer * layer, - struct ggml_tensor * cur, - struct ggml_tensor * x_prev, - struct ggml_tensor ** wkv_state, - size_t wkv_head_size, - size_t head_count_kv) { - size_t n_embd = cur->ne[0]; - size_t n_seq_tokens = cur->ne[1]; - size_t n_seqs = cur->ne[2]; - - size_t head_size = wkv_head_size; - size_t head_count = n_embd / head_size; - - size_t n_tokens = n_seqs * n_seq_tokens; - - bool is_qrwkv = layer->time_mix_first == nullptr; - - struct ggml_tensor * sx = ggml_sub(ctx, x_prev, cur); - - sx = ggml_reshape_2d(ctx, sx, n_embd, n_tokens); - cur = ggml_reshape_2d(ctx, cur, n_embd, n_tokens); - - struct ggml_tensor * xxx = ggml_add(ctx, ggml_mul(ctx, sx, layer->time_mix_lerp_x), cur); - - xxx = ggml_reshape_4d( - ctx, - ggml_tanh( - ctx, - ggml_mul_mat(ctx, layer->time_mix_w1, xxx) - ), - layer->time_mix_w1->ne[1] / 5, 1, 5, n_tokens - ); - - xxx = ggml_cont(ctx, ggml_permute(ctx, xxx, 0, 1, 3, 2)); - - xxx = ggml_mul_mat( - ctx, - ggml_reshape_4d( - ctx, - layer->time_mix_w2, - layer->time_mix_w2->ne[0], layer->time_mix_w2->ne[1], 1, 5 - ), - xxx - ); - - struct ggml_tensor *xw, *xk, *xv, *xr, *xg; - if (layer->time_mix_lerp_fused) { - // fusing these weights makes some performance improvement - sx = ggml_reshape_3d(ctx, sx, n_embd, 1, n_tokens); - cur = ggml_reshape_3d(ctx, cur, n_embd, 1, n_tokens); - xxx = ggml_add(ctx, ggml_mul(ctx, ggml_add(ctx, xxx, layer->time_mix_lerp_fused), sx), cur); - xw = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], 0); - xk = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float)); - xv = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float)); - xr = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float)); - xg = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float)); - } else { - // for backward compatibility - xw = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], 0); - xk = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float)); - xv = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float)); - xr = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float)); - xg = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float)); - - xw = ggml_add(ctx, ggml_mul(ctx, ggml_add(ctx, xw, layer->time_mix_lerp_w), sx), cur); - xk = ggml_add(ctx, ggml_mul(ctx, ggml_add(ctx, xk, layer->time_mix_lerp_k), sx), cur); - xv = ggml_add(ctx, ggml_mul(ctx, ggml_add(ctx, xv, layer->time_mix_lerp_v), sx), cur); - xr = ggml_add(ctx, ggml_mul(ctx, ggml_add(ctx, xr, layer->time_mix_lerp_r), sx), cur); - xg = ggml_add(ctx, ggml_mul(ctx, ggml_add(ctx, xg, layer->time_mix_lerp_g), sx), cur); - } - - struct ggml_tensor * r = llm_build_lora_mm(lctx, ctx, layer->time_mix_receptance, xr); - struct ggml_tensor * k = llm_build_lora_mm(lctx, ctx, layer->time_mix_key, xk); - struct ggml_tensor * v = llm_build_lora_mm(lctx, ctx, layer->time_mix_value, xv); - if (layer->time_mix_receptance_b) { - r = ggml_add(ctx, r, layer->time_mix_receptance_b); - } - if (layer->time_mix_key_b) { - k = ggml_add(ctx, k, layer->time_mix_key_b); - } - if (layer->time_mix_value_b) { - v = ggml_add(ctx, v, layer->time_mix_value_b); - } - - struct ggml_tensor * g = llm_build_lora_mm(lctx, ctx, layer->time_mix_gate, xg); - if (is_qrwkv) { - g = ggml_sigmoid(ctx, g); - } else { - g = ggml_silu(ctx, g); - } - - if (head_count_kv != head_count) { - GGML_ASSERT(head_count % head_count_kv == 0); - k = ggml_reshape_4d(ctx, k, head_size, 1, head_count_kv, n_tokens); - v = ggml_reshape_4d(ctx, v, head_size, 1, head_count_kv, n_tokens); - struct ggml_tensor * tmp = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, head_size, head_count / head_count_kv, head_count_kv, n_tokens); - k = ggml_repeat(ctx, k, tmp); - v = ggml_repeat(ctx, v, tmp); - } - - k = ggml_reshape_3d(ctx, k, head_size, head_count, n_tokens); - v = ggml_reshape_3d(ctx, v, head_size, head_count, n_tokens); - r = ggml_reshape_3d(ctx, r, head_size, head_count, n_tokens); - - struct ggml_tensor * w = ggml_mul_mat( - ctx, - layer->time_mix_decay_w2, - ggml_tanh( - ctx, - ggml_mul_mat(ctx, layer->time_mix_decay_w1, xw) - ) - ); - - w = ggml_add(ctx, w, layer->time_mix_decay); - w = ggml_exp(ctx, ggml_neg(ctx, ggml_exp(ctx, w))); - w = ggml_reshape_3d(ctx, w, head_size, head_count, n_tokens); - - if (is_qrwkv) { - // k = k * (1 - w) - k = ggml_sub(ctx, k, ggml_mul(ctx, k, w)); - } - - struct ggml_tensor * wkv_output; - if (!layer->time_mix_first) { - wkv_output = ggml_gated_linear_attn(ctx, k, v, r, w, *wkv_state, pow(head_size, -0.5f)); - } else { - wkv_output = ggml_rwkv_wkv6(ctx, k, v, r, layer->time_mix_first, w, *wkv_state); - } - cur = ggml_view_1d(ctx, wkv_output, n_embd * n_tokens, 0); - *wkv_state = ggml_view_1d(ctx, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float)); - - if (!is_qrwkv) { - // group norm with head_count groups - cur = ggml_reshape_3d(ctx, cur, n_embd / head_count, head_count, n_tokens); - cur = ggml_norm(ctx, cur, 64e-5f); - - // Convert back to regular vectors. - cur = ggml_reshape_2d(ctx, cur, n_embd, n_tokens); - cur = ggml_add(ctx, ggml_mul(ctx, cur, layer->time_mix_ln), layer->time_mix_ln_b); - } else { - cur = ggml_reshape_2d(ctx, cur, n_embd, n_tokens); - } - - cur = ggml_mul(ctx, cur, g); - cur = llm_build_lora_mm(lctx, ctx, layer->time_mix_output, cur); - - return ggml_reshape_3d(ctx, cur, n_embd, n_seq_tokens, n_seqs); -} - -static struct ggml_tensor * llm_build_rwkv6_channel_mix( - struct llama_context & lctx, - struct ggml_context * ctx, - const struct llama_layer * layer, - struct ggml_tensor * cur, - struct ggml_tensor * x_prev) { - struct ggml_tensor * sx = ggml_sub(ctx, x_prev, cur); - struct ggml_tensor * xk = ggml_add(ctx, ggml_mul(ctx, sx, layer->channel_mix_lerp_k), cur); - struct ggml_tensor * xr = ggml_add(ctx, ggml_mul(ctx, sx, layer->channel_mix_lerp_r), cur); - - struct ggml_tensor * r = ggml_sigmoid(ctx, llm_build_lora_mm(lctx, ctx, layer->channel_mix_receptance, xr)); - struct ggml_tensor * k = ggml_sqr( - ctx, - ggml_relu( - ctx, - llm_build_lora_mm(lctx, ctx, layer->channel_mix_key, xk) - ) - ); - - return ggml_mul(ctx, r, llm_build_lora_mm(lctx, ctx, layer->channel_mix_value, k)); -} - -struct llm_build_context { - const llama_model & model; - llama_context & lctx; - const llama_hparams & hparams; - const llama_cparams & cparams; - const llama_ubatch & ubatch; - const llama_kv_cache & kv_self; - - const int64_t n_embd; - const int64_t n_layer; - const int64_t n_rot; - const int64_t n_ctx; // user-specified context size (can be different from n_ctx_train) - const int64_t n_head; - const int64_t n_head_kv; - const int64_t n_embd_head_k; - const int64_t n_embd_k_gqa; - const int64_t n_embd_head_v; - const int64_t n_embd_v_gqa; - const int64_t n_expert; - const int64_t n_expert_used; - - const float freq_base; - const float freq_scale; - const float ext_factor; - const float attn_factor; - const float beta_fast; - const float beta_slow; - const float norm_eps; - const float norm_rms_eps; - - const int32_t n_tokens; - const int32_t n_kv; // size of KV cache to consider (n_kv <= kv_self.size) - const int32_t n_outputs; - const int32_t n_outputs_enc; - const int32_t kv_head; // index of where we store new KV data in the cache - const int32_t n_ctx_orig; - - const bool flash_attn; - - const enum llama_pooling_type pooling_type; - const enum llama_rope_type rope_type; - - const llm_build_cb & cb; - - std::vector & buf_compute_meta; - - struct ggml_context * ctx0 = nullptr; - - // TODO: consider making the entire interface noexcept - llm_build_context( - llama_context & lctx, - const llama_ubatch & ubatch, - const llm_build_cb & cb, - bool worst_case) : - model (lctx.model), - lctx (lctx), - hparams (model.hparams), - cparams (lctx.cparams), - ubatch (ubatch), - kv_self (lctx.kv_self), - n_embd (hparams.n_embd), - n_layer (hparams.n_layer), - n_rot (hparams.n_rot), - n_ctx (cparams.n_ctx), - n_head (hparams.n_head()), - n_head_kv (hparams.n_head_kv()), - n_embd_head_k (hparams.n_embd_head_k), - n_embd_k_gqa (hparams.n_embd_k_gqa()), - n_embd_head_v (hparams.n_embd_head_v), - n_embd_v_gqa (hparams.n_embd_v_gqa()), - n_expert (hparams.n_expert), - n_expert_used (hparams.n_expert_used), - freq_base (cparams.rope_freq_base), - freq_scale (cparams.rope_freq_scale), - ext_factor (cparams.yarn_ext_factor), - attn_factor (cparams.yarn_attn_factor), - beta_fast (cparams.yarn_beta_fast), - beta_slow (cparams.yarn_beta_slow), - norm_eps (hparams.f_norm_eps), - norm_rms_eps (hparams.f_norm_rms_eps), - n_tokens (ubatch.n_tokens), - n_kv (worst_case ? kv_self.size : kv_self.n), - n_outputs (worst_case ? n_tokens : lctx.n_outputs), - n_outputs_enc (worst_case ? n_tokens : lctx.embd_enc.size() / hparams.n_embd), - kv_head (worst_case ? (kv_self.recurrent ? 0 : kv_self.size - n_tokens) : kv_self.head), - n_ctx_orig (cparams.n_ctx_orig_yarn), - flash_attn (cparams.flash_attn), - pooling_type (cparams.pooling_type), - rope_type (hparams.rope_type), - cb (cb), - buf_compute_meta (lctx.buf_compute_meta) { - // all initializations should be done in init() - } - - void init() { - struct ggml_init_params params = { - /*.mem_size =*/ buf_compute_meta.size(), - /*.mem_buffer =*/ buf_compute_meta.data(), - /*.no_alloc =*/ true, - }; - - ctx0 = ggml_init(params); - - lctx.inp_tokens = nullptr; - lctx.inp_embd = nullptr; - lctx.inp_pos = nullptr; - lctx.inp_out_ids = nullptr; - lctx.inp_KQ_mask = nullptr; - lctx.inp_KQ_mask_swa = nullptr; - lctx.inp_K_shift = nullptr; - lctx.inp_mean = nullptr; - lctx.inp_cls = nullptr; - lctx.inp_s_copy = nullptr; - lctx.inp_s_mask = nullptr; - lctx.inp_s_seq = nullptr; - lctx.inp_pos_bucket = nullptr; - lctx.inp_embd_enc = nullptr; - lctx.inp_KQ_mask_cross = nullptr; - } - - void free() { - ggml_free(ctx0); - ctx0 = nullptr; - } - - struct ggml_cgraph * build_k_shift() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - GGML_ASSERT(kv_self.size == n_ctx); - - lctx.inp_K_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_ctx); - cb(lctx.inp_K_shift, "K_shift", -1); - ggml_set_input(lctx.inp_K_shift); - - for (int il = 0; il < n_layer; ++il) { - const int64_t n_head_kv = hparams.n_head_kv(il); - const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); - struct ggml_tensor * rope_factors = build_rope_factors(il); - struct ggml_tensor * k = - ggml_view_3d(ctx0, kv_self.k_l[il], - n_embd_head_k, n_head_kv, n_ctx, - ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k), - ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa), - 0); - - struct ggml_tensor * tmp; - if (ggml_is_quantized(k->type)) { - // dequantize to f32 -> RoPE -> quantize back - tmp = ggml_cast(ctx0, k, GGML_TYPE_F32); - cb(tmp, "K_f32", il); - for (auto & backend : lctx.backends) { - // Figure out which backend KV cache belongs to - if (ggml_backend_supports_buft(backend.get(), ggml_backend_buffer_get_type(kv_self.k_l[il]->buffer))) { - ggml_backend_sched_set_tensor_backend(lctx.sched.get(), tmp, backend.get()); - break; - } - } - tmp = ggml_rope_ext_inplace(ctx0, tmp, - lctx.inp_K_shift, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - cb(tmp, "K_shifted_f32", il); - tmp = ggml_cpy(ctx0, tmp, k); - } else { - // we rotate only the first n_rot dimensions - tmp = ggml_rope_ext_inplace(ctx0, k, - lctx.inp_K_shift, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - } - cb(tmp, "K_shifted", il); - ggml_build_forward_expand(gf, tmp); - } - - return gf; - } - - struct ggml_cgraph * build_defrag(const std::vector & ids) { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - for (uint32_t i = 0; i < ids.size(); ++i) { - const uint32_t id = ids[i]; - - if (i == id || id == ids.size()) { - continue; - } - - uint32_t nm = 1; - - while (i + nm < ids.size() && ids[i + nm] == id + nm) { - nm++; - } - - for (int il = 0; il < n_layer; ++il) { - const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); - const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); - - ggml_tensor * view_k_src = ggml_view_2d(ctx0, kv_self.k_l[il], - n_embd_k_gqa, nm, - ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa), - ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*i)); - - ggml_tensor * view_k_dst = ggml_view_2d(ctx0, kv_self.k_l[il], - n_embd_k_gqa, nm, - ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa), - ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*id)); - - ggml_tensor * view_v_src; - ggml_tensor * view_v_dst; - - if (flash_attn) { - // NOTE: the V cache is not transposed when using flash attention - view_v_src = ggml_view_2d(ctx0, kv_self.v_l[il], - n_embd_v_gqa, nm, - ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa), - ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*i)); - - view_v_dst = ggml_view_2d(ctx0, kv_self.v_l[il], - n_embd_v_gqa, nm, - ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa), - ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*id)); - } else { - view_v_src = ggml_view_2d(ctx0, kv_self.v_l[il], - nm, n_embd_v_gqa, - ggml_row_size(kv_self.v_l[il]->type, kv_self.size), - ggml_row_size(kv_self.v_l[il]->type, i)); - - view_v_dst = ggml_view_2d(ctx0, kv_self.v_l[il], - nm, n_embd_v_gqa, - ggml_row_size(kv_self.v_l[il]->type, kv_self.size), - ggml_row_size(kv_self.v_l[il]->type, id)); - } - - ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_k_src, view_k_dst)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_v_src, view_v_dst)); - } - - i += nm - 1; - } - - //LLAMA_LOG_INFO("gf->n_nodes = %d\n", gf->n_nodes); - - return gf; - } - - struct ggml_tensor * build_inp_pos() { - lctx.inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - cb(lctx.inp_pos, "inp_pos", -1); - ggml_set_input(lctx.inp_pos); - return lctx.inp_pos; - } - - struct ggml_tensor * build_rope_factors(int il) { - // choose long/short freq factors based on the context size - const auto n_ctx_pre_seq = cparams.n_ctx / cparams.n_seq_max; - - if (model.layers[il].rope_freqs != nullptr) { - return model.layers[il].rope_freqs; - } - - if (n_ctx_pre_seq > hparams.n_ctx_orig_yarn) { - return model.layers[il].rope_long; - } - - return model.layers[il].rope_short; - } - - struct ggml_tensor * build_inp_out_ids() { - lctx.inp_out_ids = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_outputs); - cb(lctx.inp_out_ids, "inp_out_ids", -1); - ggml_set_input(lctx.inp_out_ids); - return lctx.inp_out_ids; - } - - struct ggml_tensor * build_inp_KQ_mask(bool causal = true) { - lctx.inp_KQ_mask = causal - ? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)) - : ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); - cb(lctx.inp_KQ_mask, "KQ_mask", -1); - ggml_set_input(lctx.inp_KQ_mask); - - return flash_attn ? ggml_cast(ctx0, lctx.inp_KQ_mask, GGML_TYPE_F16) : lctx.inp_KQ_mask; - } - - struct ggml_tensor * build_inp_KQ_mask_swa(bool causal = true) { - GGML_ASSERT(hparams.n_swa > 0); - - lctx.inp_KQ_mask_swa = causal - ? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)) - : ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); - cb(lctx.inp_KQ_mask_swa, "KQ_mask_swa", -1); - ggml_set_input(lctx.inp_KQ_mask_swa); - - return flash_attn ? ggml_cast(ctx0, lctx.inp_KQ_mask_swa, GGML_TYPE_F16) : lctx.inp_KQ_mask_swa; - } - - struct ggml_tensor * build_inp_mean() { - lctx.inp_mean = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens); - cb(lctx.inp_mean, "inp_mean", -1); - ggml_set_input(lctx.inp_mean); - return lctx.inp_mean; - } - - struct ggml_tensor * build_inp_cls() { - lctx.inp_cls = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - cb(lctx.inp_cls, "inp_cls", -1); - ggml_set_input(lctx.inp_cls); - return lctx.inp_cls; - } - - struct ggml_tensor * build_inp_s_copy() { - lctx.inp_s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_kv); - cb(lctx.inp_s_copy, "inp_s_copy", -1); - ggml_set_input(lctx.inp_s_copy); - return lctx.inp_s_copy; - } - - struct ggml_tensor * build_inp_s_mask() { - lctx.inp_s_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, n_kv); - cb(lctx.inp_s_mask, "inp_s_mask", -1); - ggml_set_input(lctx.inp_s_mask); - return lctx.inp_s_mask; - } - - struct ggml_cgraph * append_pooling(struct ggml_cgraph * gf) { - // find result_norm tensor for input - struct ggml_tensor * inp = nullptr; - for (int i = ggml_graph_n_nodes(gf) - 1; i >= 0; --i) { - inp = ggml_graph_node(gf, i); - if (strcmp(inp->name, "result_norm") == 0 || strcmp(inp->name, "result_embd") == 0) { - break; - } else { - inp = nullptr; - } - } - GGML_ASSERT(inp != nullptr && "missing result_norm/result_embd tensor"); - - struct ggml_tensor * cur; - - switch (pooling_type) { - case LLAMA_POOLING_TYPE_NONE: - { - cur = inp; - } break; - case LLAMA_POOLING_TYPE_MEAN: - { - struct ggml_tensor * inp_mean = build_inp_mean(); - cur = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, inp)), inp_mean); - } break; - case LLAMA_POOLING_TYPE_CLS: - case LLAMA_POOLING_TYPE_LAST: - { - struct ggml_tensor * inp_cls = build_inp_cls(); - cur = ggml_get_rows(ctx0, inp, inp_cls); - } break; - case LLAMA_POOLING_TYPE_RANK: - { - struct ggml_tensor * inp_cls = build_inp_cls(); - inp = ggml_get_rows(ctx0, inp, inp_cls); - - // classification head - // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566 - GGML_ASSERT(model.cls != nullptr); - GGML_ASSERT(model.cls_b != nullptr); - - cur = ggml_add (ctx0, ggml_mul_mat(ctx0, model.cls, inp), model.cls_b); - cur = ggml_tanh(ctx0, cur); - - // some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en - // https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896 - if (model.cls_out) { - GGML_ASSERT(model.cls_out_b != nullptr); - - cur = ggml_add (ctx0, ggml_mul_mat(ctx0, model.cls_out, cur), model.cls_out_b); - } - } break; - default: - { - GGML_ABORT("unknown pooling type"); - } - } - - cb(cur, "result_embd_pooled", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_tensor * llm_build_pos_bucket(bool causal) { - if (causal) { - lctx.inp_pos_bucket = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_kv, n_tokens); - } else { - lctx.inp_pos_bucket = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_tokens, n_tokens); - } - - ggml_set_input(lctx.inp_pos_bucket); - cb(lctx.inp_pos_bucket, "pos_bucket", -1); - - return lctx.inp_pos_bucket; - } - - struct ggml_tensor * llm_build_pos_bias(struct ggml_tensor * pos_bucket, struct ggml_tensor * attn_rel_b) { - struct ggml_tensor * pos_bucket_1d = ggml_view_1d(ctx0, pos_bucket, pos_bucket->ne[0] * pos_bucket->ne[1], 0); - cb(pos_bucket_1d, "pos_bucket_1d", -1); - - struct ggml_tensor * pos_bias = ggml_get_rows(ctx0, attn_rel_b, pos_bucket_1d); - cb(pos_bias, "pos_bias", -1); - - pos_bias = ggml_view_3d(ctx0, pos_bias, pos_bias->ne[0], lctx.inp_pos_bucket->ne[0], lctx.inp_pos_bucket->ne[1], ggml_element_size(pos_bias) * pos_bias->ne[0], ggml_element_size(pos_bias) * pos_bias->ne[0] * lctx.inp_pos_bucket->ne[0], 0); - cb(pos_bias, "pos_bias", -1); - - pos_bias = ggml_permute(ctx0, pos_bias, 2, 0, 1, 3); - cb(pos_bias, "pos_bias", -1); - - pos_bias = ggml_cont(ctx0, pos_bias); - cb(pos_bias, "pos_bias", -1); - - return pos_bias; - } - - struct ggml_tensor * llm_build_inp_embd_enc() { - const int64_t n_embd = hparams.n_embd; - lctx.inp_embd_enc = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_outputs_enc); - ggml_set_input(lctx.inp_embd_enc); - cb(lctx.inp_embd_enc, "embd_enc", -1); - return lctx.inp_embd_enc; - } - - struct ggml_tensor * llm_build_inp_KQ_mask_cross() { - lctx.inp_KQ_mask_cross = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_outputs_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); - ggml_set_input(lctx.inp_KQ_mask_cross); - cb(lctx.inp_KQ_mask_cross, "KQ_mask_cross", -1); - return lctx.inp_KQ_mask_cross; - } - - struct ggml_cgraph * build_llama() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - // mutable variable, needed during the last layer of the computation to skip unused tokens - int32_t n_tokens = this->n_tokens; - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // rope freq factors for llama3; may return nullptr for llama2 and other models - struct ggml_tensor * rope_factors = build_rope_factors(il); - - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - n_tokens = n_outputs; - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - // For Granite architecture - if (hparams.f_residual_scale) { - cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - if (model.layers[il].ffn_gate_inp == nullptr) { - - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } else { - // MoE branch - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_moe_ffn(ctx0, lctx, cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, true, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - cb, il); - cb(cur, "ffn_moe_out", il); - } - - // For Granite architecture - if (hparams.f_residual_scale) { - cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - - // For Granite architecture - if (hparams.f_logit_scale) { - cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale); - } - - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_deci() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - // mutable variable, needed during the last layer of the computation to skip unused tokens - int32_t n_tokens = this->n_tokens; - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - const int64_t n_head_kv = hparams.n_head_kv(il); - const int64_t n_head = hparams.n_head(il); - - if (n_head == 0) { - // attention-free layer of Llama-3_1-Nemotron-51B - cur = inpL; - } else { - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - } - - if (n_head > 0 && n_head_kv == 0) { - // "linear attention" of Llama-3_1-Nemotron-51B - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo, cur); - cb(cur, "wo", il); - } else if (n_head > 0) { - // self-attention - // rope freq factors for llama3; may return nullptr for llama2 and other models - struct ggml_tensor * rope_factors = build_rope_factors(il); - - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - n_tokens = n_outputs; - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - // For Granite architecture - if (hparams.f_residual_scale) { - cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); - } - - // modified to support attention-free layer of Llama-3_1-Nemotron-51B - struct ggml_tensor * ffn_inp = cur; - if (n_head > 0) { - ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - } - - // feed-forward network - if (model.layers[il].ffn_gate_inp == nullptr) { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } - - // For Granite architecture - if (hparams.f_residual_scale) { - cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - - // For Granite architecture - if (hparams.f_logit_scale) { - cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale); - } - - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_baichuan() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr; - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - switch (model.type) { - case LLM_TYPE_7B: - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - break; - case LLM_TYPE_13B: - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd/n_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd/n_head, n_head, n_tokens); - break; - default: - GGML_ABORT("fatal error"); - } - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_xverse() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, NULL, LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_falcon() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * attn_norm; - - attn_norm = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, cb, il); - cb(attn_norm, "attn_norm", il); - - // self-attention - { - if (model.layers[il].attn_norm_2) { - // Falcon-40B - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm_2, - model.layers[il].attn_norm_2_b, - LLM_NORM, cb, il); - cb(cur, "attn_norm_2", il); - } else { - cur = attn_norm; - } - - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - - // using mode = 2 for neox mode - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, - freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, - freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - attn_norm = ggml_get_rows(ctx0, attn_norm, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = cur; - - // feed forward - { - cur = llm_build_ffn(ctx0, lctx, attn_norm, // !! use the attn norm, not the result - model.layers[il].ffn_up, NULL, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = ggml_add(ctx0, cur, inpL); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - // norm - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, - model.output_norm_b, - LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_grok() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - // mutable variable, needed during the last layer of the computation to skip unused tokens - int32_t n_tokens = this->n_tokens; - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // multiply by embedding_multiplier_scale of 78.38367176906169 - inpL = ggml_scale(ctx0, inpL, 78.38367176906169f); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - n_tokens = n_outputs; - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - // Grok - // if attn_out_norm is present then apply it before adding the input - if (model.layers[il].attn_out_norm) { - cur = llm_build_norm(ctx0, cur, hparams, - model.layers[il].attn_out_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_out_norm", il); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - // MoE branch - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_moe_ffn(ctx0, lctx, cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_GELU, true, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - cb, il); - cb(cur, "ffn_moe_out", il); - - // Grok - // if layer_out_norm is present then apply it before adding the input - // Idea: maybe ffn_out_norm is a better name - if (model.layers[il].layer_out_norm) { - cur = llm_build_norm(ctx0, cur, hparams, - model.layers[il].layer_out_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "layer_out_norm", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - - // Grok - // multiply logits by output_multiplier_scale of 0.5773502691896257 - - cur = ggml_scale(ctx0, cur, 0.5773502691896257f); - - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_dbrx() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - // mutable variable, needed during the last layer of the computation to skip unused tokens - int32_t n_tokens = this->n_tokens; - - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - struct ggml_tensor * Qcur = nullptr; - struct ggml_tensor * Kcur = nullptr; - struct ggml_tensor * Vcur = nullptr; - - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); - cb(cur, "wqkv_clamped", il); - - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - n_tokens = n_outputs; - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - // MoE branch - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].attn_out_norm, NULL, - LLM_NORM, cb, il); - cb(cur, "attn_out_norm", il); - - cur = llm_build_moe_ffn(ctx0, lctx, cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, true, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - cb, il); - cb(cur, "ffn_moe_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_starcoder() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - struct ggml_tensor * pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); - cb(pos, "pos_embd", -1); - - inpL = ggml_add(ctx0, inpL, pos); - cb(inpL, "inpL", -1); - - for (int il = 0; il < n_layer; ++il) { - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - - struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // add the input - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - // FF - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = llm_build_norm(ctx0, inpL, hparams, - model.output_norm, - model.output_norm_b, - LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_refact() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - cb(Kcur, "Kcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - cb(Qcur, "Qcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_bert() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - struct ggml_tensor * inp_pos = nullptr; - - if (model.arch != LLM_ARCH_JINA_BERT_V2) { - inp_pos = build_inp_pos(); - } - - // construct input embeddings (token, type, position) - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // token types are hardcoded to zero ("Sentence A") - struct ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0); - inpL = ggml_add(ctx0, inpL, type_row0); - if (model.arch == LLM_ARCH_BERT) { - inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL); - } - cb(inpL, "inp_embd", -1); - - // embed layer norm - inpL = llm_build_norm(ctx0, inpL, hparams, model.tok_norm, model.tok_norm_b, LLM_NORM, cb, -1); - cb(inpL, "inp_norm", -1); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(false); - - // iterate layers - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * cur = inpL; - - struct ggml_tensor * Qcur; - struct ggml_tensor * Kcur; - struct ggml_tensor * Vcur; - - // self-attention - if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_JINA_BERT_V2) { - Qcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur), model.layers[il].bq); - cb(Qcur, "Qcur", il); - - if (model.layers[il].attn_q_norm) { - Qcur = llm_build_norm(ctx0, Qcur, hparams, - model.layers[il].attn_q_norm, - model.layers[il].attn_q_norm_b, - LLM_NORM, cb, il); - } - - Kcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur), model.layers[il].bk); - cb(Kcur, "Kcur", il); - - if (model.layers[il].attn_k_norm) { - Kcur = llm_build_norm(ctx0, Kcur, hparams, - model.layers[il].attn_k_norm, - model.layers[il].attn_k_norm_b, - LLM_NORM, cb, il); - } - Vcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur), model.layers[il].bv); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - } else { - // compute Q and K and RoPE them - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - } - - struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3)); - - struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); - cb(kq, "kq", il); - - kq = ggml_soft_max_ext(ctx0, kq, KQ_mask, 1.0f/sqrtf(float(n_embd_head)), hparams.f_max_alibi_bias); - cb(kq, "kq_soft_max_ext", il); - - struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens))); - cb(v, "v", il); - - struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_tokens, n_embd_head, n_head_kv), kq); - cb(kqv, "kqv", il); - - struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); - cb(kqv_merged, "kqv_merged", il); - - cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); - cb(cur, "kqv_merged_cont", il); - - ggml_build_forward_expand(gf, cur); - - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo, cur); - if (model.layers[il].bo) { - cb(cur, "kqv_wo", il); - } - - if (model.layers[il].bo) { - cur = ggml_add(ctx0, cur, model.layers[il].bo); - } - cb(cur, "kqv_out", il); - - if (il == n_layer - 1 && pooling_type == LLAMA_POOLING_TYPE_NONE) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // re-add the layer input - cur = ggml_add(ctx0, cur, inpL); - - // attention layer norm - cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, cb, il); - - if (model.layers[il].attn_norm_2 != nullptr) { - cur = ggml_add(ctx0, cur, inpL); // re-add the layer input - cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, cb, il); - } - - struct ggml_tensor * ffn_inp = cur; - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - if (model.arch == LLM_ARCH_BERT) { - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); - } else if (model.arch == LLM_ARCH_JINA_BERT_V2) { - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_PAR, cb, il); - } else { - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - } - cb(cur, "ffn_out", il); - - // attentions bypass the intermediate layer - cur = ggml_add(ctx0, cur, ffn_inp); - - // output layer norm - cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].layer_out_norm, model.layers[il].layer_out_norm_b, LLM_NORM, cb, il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cb(cur, "result_embd", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_bloom() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - inpL = llm_build_norm(ctx0, inpL, hparams, - model.tok_norm, - model.tok_norm_b, - LLM_NORM, cb, -1); - cb(inpL, "inp_norm", -1); - - for (int il = 0; il < n_layer; ++il) { - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - - struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // Add the input - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - // FF - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = llm_build_norm(ctx0, inpL, hparams, - model.output_norm, - model.output_norm_b, - LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_mpt() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - struct ggml_tensor * cur; - struct ggml_tensor * pos; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - if (model.pos_embd) { - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); - cb(pos, "pos_embd", -1); - - inpL = ggml_add(ctx0, inpL, pos); - cb(inpL, "inpL", -1); - } - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * attn_norm; - - attn_norm = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, cb, il); - cb(attn_norm, "attn_norm", il); - - // self-attention - { - cur = attn_norm; - - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - if (model.layers[il].bqkv){ - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - } - - if (hparams.f_clamp_kqv > 0.0f) { - cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); - cb(cur, "wqkv_clamped", il); - } - - struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - // Q/K Layernorm - if (model.layers[il].attn_q_norm) { - Qcur = llm_build_norm(ctx0, Qcur, hparams, - model.layers[il].attn_q_norm, - model.layers[il].attn_q_norm_b, - LLM_NORM, cb, il); - cb(Qcur, "Qcur", il); - - Kcur = llm_build_norm(ctx0, Kcur, hparams, - model.layers[il].attn_k_norm, - model.layers[il].attn_k_norm_b, - LLM_NORM, cb, il); - cb(Kcur, "Kcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } else { - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // Add the input - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - // feed forward - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, cb, il); - cb(cur, "ffn_norm", il); - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - model.layers[il].ffn_act, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, - model.output_norm_b, - LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_stablelm() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, cb, il); - cb(cur, "attn_norm", il); - - struct ggml_tensor * inpSA = cur; - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - cb(Qcur, "Qcur", il); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - cb(Kcur, "Kcur", il); - - if (model.layers[il].attn_q_norm) { - Qcur = llm_build_norm(ctx0, Qcur, hparams, - model.layers[il].attn_q_norm, - NULL, - LLM_NORM, cb, il); - cb(Qcur, "Qcur", il); - } - if (model.layers[il].attn_k_norm) { - Kcur = llm_build_norm(ctx0, Kcur, hparams, - model.layers[il].attn_k_norm, - NULL, - LLM_NORM, cb, il); - cb(Kcur, "Kcur", il); - } - - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - if (model.layers[il].ffn_norm) { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, cb, il); - cb(cur, "ffn_norm", il); - } else { - // parallel residual - cur = inpSA; - } - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, - model.output_norm_b, - LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_qwen() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - - struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 2*sizeof(float)*(n_embd))); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - - // using mode = 2 for neox mode - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, - freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, - freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward forward - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_qwen2() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_qwen2vl() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - lctx.inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens * 4); - cb(lctx.inp_pos, "inp_pos", -1); - ggml_set_input(lctx.inp_pos); - struct ggml_tensor * inp_pos = lctx.inp_pos; - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - int sections[4]; - std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - - Qcur = ggml_rope_multi( - ctx0, - ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_multi( - ctx0, - ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_qwen2moe() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - // mutable variable, needed during the last layer of the computation to skip unused tokens - int32_t n_tokens = this->n_tokens; - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self_attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - n_tokens = n_outputs; - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // MoE branch - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - ggml_tensor * moe_out = - llm_build_moe_ffn(ctx0, lctx, cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, false, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - cb, il); - cb(cur, "ffn_moe_out", il); - - // FFN shared expert - { - ggml_tensor * cur_gate_inp = llm_build_lora_mm(lctx, ctx0, model.layers[il].ffn_gate_inp_shexp, cur); - cb(cur_gate_inp, "ffn_shexp_gate_inp", il); - - // sigmoid - ggml_tensor * cur_gate = ggml_div(ctx0, ggml_silu(ctx0, cur_gate_inp), cur_gate_inp); - cb(cur_gate, "ffn_shexp_gate", il); - - ggml_tensor * cur_ffn = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up_shexp, NULL, NULL, - model.layers[il].ffn_gate_shexp, NULL, NULL, - model.layers[il].ffn_down_shexp, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur_ffn, "ffn_shexp", il); - - ggml_tensor * ffn_shexp_out = ggml_mul(ctx0, cur_ffn, cur_gate); - cb(ffn_shexp_out, "ffn_shexp_out", il); - - moe_out = ggml_add(ctx0, moe_out, ffn_shexp_out); - cb(moe_out, "ffn_out", il); - - cur = moe_out; - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_phi2() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - struct ggml_tensor * cur; - struct ggml_tensor * attn_norm_output; - struct ggml_tensor * ffn_output; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - attn_norm_output = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, cb, il); - cb(attn_norm_output, "attn_norm", il); - - // self-attention - { - struct ggml_tensor * Qcur = nullptr; - struct ggml_tensor * Kcur = nullptr; - struct ggml_tensor * Vcur = nullptr; - - if (model.layers[il].wqkv) { - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, attn_norm_output); - cb(cur, "wqkv", il); - - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - } else { - Qcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, attn_norm_output), model.layers[il].bq); - Kcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, attn_norm_output), model.layers[il].bk); - Vcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, attn_norm_output), model.layers[il].bv); - } - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, - freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - // with phi2, we scale the Q to avoid precision issues - // ref: https://github.com/ml-explore/mlx-examples/blob/08e862336ade809bc37d1035f94b359e7d1a5152/phi2/phi2.py#L64-L66 - Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head))); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, - freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids); - } - - // FF - { - ffn_output = llm_build_ffn(ctx0, lctx, attn_norm_output, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); - cb(ffn_output, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_output); - cur = ggml_add(ctx0, cur, inpL); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = llm_build_norm(ctx0, inpL, hparams, - model.output_norm, - model.output_norm_b, - LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output_no_bias", -1); - - cur = ggml_add(ctx0, cur, model.output_b); - cb(cur, "result_output", -1); - ggml_build_forward_expand(gf, cur); - return gf; - } - - struct ggml_cgraph * build_phi3() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = nullptr; - if (hparams.n_swa == 0) { - // Phi-4 doesn't use sliding window attention - KQ_mask = build_inp_KQ_mask(); - } else { - KQ_mask = build_inp_KQ_mask_swa(); - } - - for (int il = 0; il < n_layer; ++il) { - auto residual = inpL; - - // self-attention - { - // rope freq factors for 128k context - struct ggml_tensor * rope_factors = build_rope_factors(il); - - struct ggml_tensor* attn_norm_output = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM_RMS, cb, il); - cb(attn_norm_output, "attn_norm", il); - - struct ggml_tensor * Qcur = nullptr; - struct ggml_tensor * Kcur = nullptr; - struct ggml_tensor * Vcur = nullptr; - - if (model.layers[il].wqkv) { - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, attn_norm_output); - cb(cur, "wqkv", il); - - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0 * sizeof(float) * (n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd))); - Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa))); - } else { - Qcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, attn_norm_output), model.layers[il].bq); - Kcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, attn_norm_output), model.layers[il].bk); - Vcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, attn_norm_output), model.layers[il].bv); - } - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, - freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head))); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, - freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor* inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - residual = ggml_get_rows(ctx0, residual, inp_out_ids); - } - - cur = ggml_add(ctx0, cur, residual); - residual = cur; - - cur = llm_build_norm(ctx0, cur, hparams, - model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - // feed-forward network - if (model.layers[il].ffn_gate_inp == nullptr) { - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SWIGLU, LLM_FFN_SEQ, cb, il); - cb(cur, "ffn_out", il); - } else { - // MoE branch - cur = llm_build_moe_ffn(ctx0, lctx, cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, true, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - cb, il); - cb(cur, "ffn_moe_out", il); - } - - cur = ggml_add(ctx0, residual, cur); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = llm_build_norm(ctx0, inpL, hparams, - model.output_norm, - model.output_norm_b, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - - if (model.output_b != nullptr) { - cb(cur, "result_output_no_bias", -1); - cur = ggml_add(ctx0, cur, model.output_b); - } - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - - struct ggml_cgraph * build_plamo() { - struct ggml_cgraph * gf = ggml_new_graph(ctx0); - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - struct ggml_tensor * attention_norm = cur; - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_rot, n_head, n_tokens), inp_pos, nullptr, - n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens), inp_pos, nullptr, - n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - struct ggml_tensor * sa_out = cur; - - cur = attention_norm; - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - sa_out = ggml_get_rows(ctx0, sa_out, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // feed-forward network - { - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, sa_out); - cur = ggml_add(ctx0, cur, inpL); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_gpt2() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - struct ggml_tensor * cur; - struct ggml_tensor * pos; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); - cb(pos, "pos_embd", -1); - - inpL = ggml_add(ctx0, inpL, pos); - cb(inpL, "inpL", -1); - - for (int il = 0; il < n_layer; ++il) { - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - - struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // add the input - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - // FF - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = llm_build_norm(ctx0, inpL, hparams, - model.output_norm, - model.output_norm_b, - LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_codeshell() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - - struct ggml_tensor * tmpq = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - struct ggml_tensor * tmpk = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - - cb(tmpq, "tmpq", il); - cb(tmpk, "tmpk", il); - cb(Vcur, "Vcur", il); - - struct ggml_tensor * Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // add the input - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - // FF - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = llm_build_norm(ctx0, inpL, hparams, - model.output_norm, - model.output_norm_b, - LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_orion() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, model.layers[il].attn_norm_b, - LLM_NORM, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - // if (model.layers[il].bq) { - // Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - // cb(Qcur, "Qcur", il); - // } - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - // if (model.layers[il].bk) { - // Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - // cb(Kcur, "Kcur", il); - // } - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - // if (model.layers[il].bv) { - // Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - // cb(Vcur, "Vcur", il); - // } - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, - LLM_NORM, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, model.output_norm_b, - LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_internlm2() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_minicpm3() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - //TODO: if the model varies, these parameters need to be read from the model - const int64_t n_embd_base = 256; - const float scale_embd = 12.0f; - const float scale_depth = 1.4f; - const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k)); - - const uint32_t n_embd_head_qk_rope = hparams.n_rot; - const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; - const uint32_t kv_lora_rank = hparams.n_lora_kv; - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // scale the input embeddings - inpL = ggml_scale(ctx0, inpL, scale_embd); - cb(inpL, "inp_scaled", -1); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - struct ggml_tensor * rope_factors = build_rope_factors(il); - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self_attention - { - struct ggml_tensor * q = NULL; - // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens} - q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur); - cb(q, "q", il); - - q = llm_build_norm(ctx0, q, hparams, - model.layers[il].attn_q_a_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(q, "q", il); - - // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens} - q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q); - cb(q, "q", il); - - // split into {n_head * n_embd_head_qk_nope, n_tokens} - struct ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, - ggml_row_size(q->type, hparams.n_embd_head_k), - ggml_row_size(q->type, hparams.n_embd_head_k * n_head), - 0); - cb(q_nope, "q_nope", il); - - // and {n_head * n_embd_head_qk_rope, n_tokens} - struct ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, - ggml_row_size(q->type, hparams.n_embd_head_k), - ggml_row_size(q->type, hparams.n_embd_head_k * n_head), - ggml_row_size(q->type, n_embd_head_qk_nope)); - cb(q_pe, "q_pe", il); - - // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens} - struct ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); - cb(kv_pe_compresseed, "kv_pe_compresseed", il); - - // split into {kv_lora_rank, n_tokens} - struct ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens, - kv_pe_compresseed->nb[1], - 0); - cb(kv_compressed, "kv_compressed", il); - - // and {n_embd_head_qk_rope, n_tokens} - struct ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens, - kv_pe_compresseed->nb[1], - kv_pe_compresseed->nb[1], - ggml_row_size(kv_pe_compresseed->type, kv_lora_rank)); - cb(k_pe, "k_pe", il); - - // TODO: the CUDA backend used to not support non-cont. (RMS) norm, investigate removing ggml_cont - kv_compressed = ggml_cont(ctx0, kv_compressed); - kv_compressed = llm_build_norm(ctx0, kv_compressed, hparams, - model.layers[il].attn_kv_a_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(kv_compressed, "kv_compressed", il); - - // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens} - struct ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed); - cb(kv, "kv", il); - - // split into {n_head * n_embd_head_qk_nope, n_tokens} - struct ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens, - ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v), - ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)), - 0); - cb(k_nope, "k_nope", il); - - // and {n_head * n_embd_head_v, n_tokens} - struct ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens, - ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)), - ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head), - ggml_row_size(kv->type, (n_embd_head_qk_nope))); - cb(v_states, "v_states", il); - - v_states = ggml_cont(ctx0, v_states); - cb(v_states, "v_states", il); - - v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens, - ggml_row_size(kv->type, hparams.n_embd_head_v * n_head), - 0); - cb(v_states, "v_states", il); - - q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this - q_pe = ggml_rope_ext( - ctx0, q_pe, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(q_pe, "q_pe", il); - - // shared RoPE key - k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this - k_pe = ggml_rope_ext( - ctx0, k_pe, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(k_pe, "k_pe", il); - - struct ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0); - cb(q_states, "q_states", il); - - struct ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0); - cb(k_states, "k_states", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - k_states, v_states, q_states, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - // scale_res - scale the hidden states for residual connection - const float scale_res = scale_depth/sqrtf(float(n_layer)); - cur = ggml_scale(ctx0, cur, scale_res); - cb(cur, "hidden_scaled", il); - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } - - // scale the hidden states for residual connection - cur = ggml_scale(ctx0, cur, scale_res); - cb(cur, "hidden_scaled_ffn", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head scaling - const float scale_lmhead = float(n_embd_base)/float(n_embd); - cur = ggml_scale(ctx0, cur, scale_lmhead); - cb(cur, "lmhead_scaling", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_gemma() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head_k = hparams.n_embd_head_k; - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); - cb(inpL, "inp_scaled", -1); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - cb(Qcur, "Qcur", il); - - Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k))); - cb(Qcur, "Qcur_scaled", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); - cb(sa_out, "sa_out", il); - - cur = llm_build_norm(ctx0, sa_out, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - // feed-forward network - { - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, sa_out); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_gemma2() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head_k = hparams.n_embd_head_k; - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); - cb(inpL, "inp_scaled", -1); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - // gemma 2 requires different mask for layers using sliding window (SWA) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(true); - struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa(true); - - for (int il = 0; il < n_layer; ++il) { - // (il % 2) layers use SWA - struct ggml_tensor * KQ_mask_l = (il % 2 == 0) ? KQ_mask_swa : KQ_mask; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - cb(Qcur, "Qcur", il); - - // ref: https://github.com/google/gemma_pytorch/commit/03e657582d17cb5a8617ebf333c1c16f3694670e - switch (model.type) { - case LLM_TYPE_2B: - case LLM_TYPE_9B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k))); break; - case LLM_TYPE_27B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd / n_head))); break; - default: GGML_ABORT("fatal error"); - }; - cb(Qcur, "Qcur_scaled", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask_l, n_tokens, kv_head, n_kv, 1.0f, cb, il); - } - - cur = llm_build_norm(ctx0, cur, hparams, - model.layers[il].attn_post_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_post_norm", il); - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); - cb(sa_out, "sa_out", il); - - cur = llm_build_norm(ctx0, sa_out, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - // feed-forward network - { - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } - - cur = llm_build_norm(ctx0, cur, hparams, - model.layers[il].ffn_post_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "ffn_post_norm", -1); - - cur = ggml_add(ctx0, cur, sa_out); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - - // final logit soft-capping - cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping); - cur = ggml_tanh(ctx0, cur); - cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping); - - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - - struct ggml_cgraph * build_starcoder2() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, model.layers[il].attn_norm_b, - LLM_NORM, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, - LLM_NORM, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, model.output_norm_b, - LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_mamba() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - // {n_embd, n_tokens} - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - struct ggml_tensor * state_copy = build_inp_s_copy(); - struct ggml_tensor * state_mask = build_inp_s_mask(); - - for (int il = 0; il < n_layer; ++il) { - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - cur = llm_build_mamba(ctx0, lctx, ubatch, gf, cur, - state_copy, state_mask, - kv_head, n_kv, cb, il); - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // residual - cur = ggml_add(ctx0, cur, inpL); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - // final rmsnorm - cur = llm_build_norm(ctx0, inpL, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_command_r() { - - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - const float f_logit_scale = hparams.f_logit_scale; - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM, cb, il); - cb(cur, "attn_norm", il); - struct ggml_tensor * ffn_inp = cur; - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - if (model.layers[il].attn_q_norm) { - Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens, - ggml_element_size(Qcur) * n_embd_head, - ggml_element_size(Qcur) * n_embd_head * n_head, - 0); - cb(Qcur, "Qcur", il); - Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens, - ggml_element_size(Kcur) * n_embd_head, - ggml_element_size(Kcur) * n_embd_head * n_head_kv, - 0); - cb(Kcur, "Kcur", il); - - Qcur = llm_build_norm(ctx0, Qcur, hparams, - model.layers[il].attn_q_norm, - NULL, - LLM_NORM, cb, il); - cb(Qcur, "Qcur", il); - - Kcur = llm_build_norm(ctx0, Kcur, hparams, - model.layers[il].attn_k_norm, - NULL, - LLM_NORM, cb, il); - cb(Kcur, "Kcur", il); - } - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); - } - - struct ggml_tensor * attn_out = cur; - - // feed-forward network - { - cur = llm_build_ffn(ctx0, lctx, ffn_inp, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } - - // add together residual + FFN + self-attention - cur = ggml_add(ctx0, cur, inpL); - cur = ggml_add(ctx0, cur, attn_out); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - - if (f_logit_scale) { - cur = ggml_scale(ctx0, cur, f_logit_scale); - } - - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - - } - - struct ggml_cgraph * build_cohere2() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - const float f_logit_scale = hparams.f_logit_scale; - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - // cohere2 requires different mask for layers using sliding window (SWA) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa(); - - // sliding window switch pattern - const int32_t sliding_window_pattern = 4; - - for (int il = 0; il < n_layer; ++il) { - // three layers sliding window attention (window size 4096) and ROPE - // fourth layer uses global attention without positional embeddings - const bool is_sliding = il % sliding_window_pattern < (sliding_window_pattern - 1); - struct ggml_tensor * KQ_mask_l = is_sliding ? KQ_mask_swa : KQ_mask; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, model.layers[il].attn_norm, NULL, LLM_NORM, cb, il); - cb(cur, "attn_norm", il); - struct ggml_tensor * ffn_inp = cur; - - // self-attention - { - // rope freq factors for 128k context - struct ggml_tensor * rope_factors = build_rope_factors(il); - - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - if (is_sliding) { - Qcur = ggml_rope_ext(ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, - beta_fast, beta_slow); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext(ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, - rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor, - attn_factor, beta_fast, beta_slow); - cb(Kcur, "Kcur", il); - } else { - // For non-sliding layers, just reshape without applying RoPE - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - cb(Qcur, "Qcur", il); - - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - cb(Kcur, "Kcur", il); - } - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, model.layers[il].wo, model.layers[il].bo, Kcur, Vcur, Qcur, - KQ_mask_l, n_tokens, kv_head, n_kv, 1.0f / sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); - } - - struct ggml_tensor * attn_out = cur; - - // feed-forward network - { - cur = llm_build_ffn(ctx0, lctx, ffn_inp, model.layers[il].ffn_up, NULL, NULL, model.layers[il].ffn_gate, - NULL, NULL, model.layers[il].ffn_down, NULL, NULL, NULL, LLM_FFN_SILU, LLM_FFN_PAR, - cb, il); - cb(cur, "ffn_out", il); - } - - // add together residual + FFN + self-attention - cur = ggml_add(ctx0, cur, inpL); - cur = ggml_add(ctx0, cur, attn_out); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, NULL, LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - - if (f_logit_scale) { - cur = ggml_scale(ctx0, cur, f_logit_scale); - } - - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - // ref: https://allenai.org/olmo - // based on the original build_llama() function, changes: - // * non-parametric layer norm - // * clamp qkv - // * removed bias - // * removed MoE - struct ggml_cgraph * build_olmo() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - // mutable variable, needed during the last layer of the computation to skip unused tokens - int32_t n_tokens = this->n_tokens; - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - NULL, NULL, - LLM_NORM, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (hparams.f_clamp_kqv > 0.0f) { - Qcur = ggml_clamp(ctx0, Qcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); - cb(Qcur, "Qcur", il); - } - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (hparams.f_clamp_kqv > 0.0f) { - Kcur = ggml_clamp(ctx0, Kcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); - cb(Kcur, "Kcur", il); - } - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (hparams.f_clamp_kqv > 0.0f) { - Vcur = ggml_clamp(ctx0, Vcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, nullptr, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - n_tokens = n_outputs; - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = llm_build_norm(ctx0, ffn_inp, hparams, - NULL, NULL, - LLM_NORM, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - NULL, NULL, - LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_olmo2() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - // mutable variable, needed during the last layer of the computation to skip unused tokens - int32_t n_tokens = this->n_tokens; - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - cur = inpL; - - // self_attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].attn_q_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(Qcur, "Qcur_normed", il); - - Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].attn_k_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(Kcur, "Kcur_normed", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur_rope", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur_rope", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - cur = llm_build_norm(ctx0, cur, hparams, - model.layers[il].attn_post_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_post_norm", il); - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - n_tokens = n_outputs; - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = llm_build_ffn(ctx0, lctx, ffn_inp, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - - cur = llm_build_norm(ctx0, cur, hparams, - model.layers[il].ffn_post_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "ffn_post_norm", -1); - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - // based on the build_qwen2moe() function, changes: - // * removed shared experts - // * removed bias - // * added q, k norm - struct ggml_cgraph * build_olmoe() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - // mutable variable, needed during the last layer of the computation to skip unused tokens - int32_t n_tokens = this->n_tokens; - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self_attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].attn_q_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(Qcur, "Qcur_normed", il); - - Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].attn_k_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(Kcur, "Kcur_normed", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur_rope", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur_rope", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - n_tokens = n_outputs; - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // MoE branch - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_moe_ffn(ctx0, lctx, cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, false, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - cb, il); - cb(cur, "ffn_moe_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_openelm() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - const int64_t n_head = hparams.n_head(il); - const int64_t n_head_kv = hparams.n_head_kv(il); - const int64_t n_head_qkv = 2*n_head_kv + n_head; - - cur = inpL; - struct ggml_tensor * residual = cur; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_reshape_3d(ctx0, cur, n_embd_head_k, n_head_qkv, n_tokens); - - struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, cur->nb[1], cur->nb[2], 0)); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*n_head)); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*(n_head+n_head_kv))); - cb(Vcur, "Vcur", il); - - Qcur = llm_build_norm(ctx0, Qcur, hparams, - model.layers[il].attn_q_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(Qcur, "Qcur", il); - - Kcur = llm_build_norm(ctx0, Kcur, hparams, - model.layers[il].attn_k_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(Kcur, "Kcur", il); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, NULL, n_rot, rope_type, n_ctx_orig, - freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, NULL, n_rot, rope_type, n_ctx_orig, - freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - Vcur = ggml_reshape_2d(ctx0, Vcur, n_embd_head * n_head_kv, n_tokens); - cb(Qcur, "Vcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - residual = ggml_get_rows(ctx0, residual, inp_out_ids); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - inpL = cur; - } - - cur = inpL; - - // norm - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_gptneox() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - - struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // ffn - if (hparams.use_par_res) { - // attention and ffn are computed in parallel - // x = x + attn(ln1(x)) + ffn(ln2(x)) - - struct ggml_tensor * attn_out = cur; - - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, inpL); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, attn_out); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } else { - // attention and ffn are computed sequentially - // x = x + attn(ln1(x)) - // x = x + ffn(ln2(x)) - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - } - - cur = llm_build_norm(ctx0, inpL, hparams, - model.output_norm, - model.output_norm_b, - LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_arctic() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - // mutable variable, needed during the last layer of the computation to skip unused tokens - int32_t n_tokens = this->n_tokens; - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - n_tokens = n_outputs; - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - - struct ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp); - cb(ffn_out, "ffn_out", il); - - // MoE - cur = llm_build_norm(ctx0, inpSA, hparams, - model.layers[il].ffn_norm_exps, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm_exps", il); - - cur = llm_build_moe_ffn(ctx0, lctx, cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, true, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - cb, il); - cb(cur, "ffn_moe_out", il); - - cur = ggml_add(ctx0, cur, ffn_out); - cb(cur, "ffn_out", il); - - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_deepseek() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - // mutable variable, needed during the last layer of the computation to skip unused tokens - int32_t n_tokens = this->n_tokens; - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // rope freq factors for llama3; may return nullptr for llama2 and other models - struct ggml_tensor * rope_factors = build_rope_factors(il); - - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - n_tokens = n_outputs; - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - if ((uint32_t) il < hparams.n_layer_dense_lead) { - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } else { - // MoE branch - ggml_tensor * moe_out = - llm_build_moe_ffn(ctx0, lctx, cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, false, - false, hparams.expert_weights_scale, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - cb, il); - cb(moe_out, "ffn_moe_out", il); - - // FFN shared expert - { - ggml_tensor * ffn_shexp = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up_shexp, NULL, NULL, - model.layers[il].ffn_gate_shexp, NULL, NULL, - model.layers[il].ffn_down_shexp, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(ffn_shexp, "ffn_shexp", il); - - cur = ggml_add(ctx0, moe_out, ffn_shexp); - cb(cur, "ffn_out", il); - } - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_deepseek2() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - // mutable variable, needed during the last layer of the computation to skip unused tokens - int32_t n_tokens = this->n_tokens; - - bool is_lite = (hparams.n_layer == 27); - - // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly. - // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation. - const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale)); - const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(hparams.n_embd_head_k)); - const float attn_factor_scaled = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale)); - - const uint32_t n_embd_head_qk_rope = hparams.n_rot; - const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; - const uint32_t kv_lora_rank = hparams.n_lora_kv; - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - // {n_embd, n_tokens} - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self_attention - { - struct ggml_tensor * q = NULL; - if (!is_lite) { - // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens} - q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur); - cb(q, "q", il); - - q = llm_build_norm(ctx0, q, hparams, - model.layers[il].attn_q_a_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(q, "q", il); - - // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens} - q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q); - cb(q, "q", il); - } else { - q = ggml_mul_mat(ctx0, model.layers[il].wq, cur); - cb(q, "q", il); - } - - // split into {n_head * n_embd_head_qk_nope, n_tokens} - struct ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, - ggml_row_size(q->type, hparams.n_embd_head_k), - ggml_row_size(q->type, hparams.n_embd_head_k * n_head), - 0); - cb(q_nope, "q_nope", il); - - // and {n_head * n_embd_head_qk_rope, n_tokens} - struct ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, - ggml_row_size(q->type, hparams.n_embd_head_k), - ggml_row_size(q->type, hparams.n_embd_head_k * n_head), - ggml_row_size(q->type, n_embd_head_qk_nope)); - cb(q_pe, "q_pe", il); - - // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens} - struct ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); - cb(kv_pe_compresseed, "kv_pe_compresseed", il); - - // split into {kv_lora_rank, n_tokens} - struct ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens, - kv_pe_compresseed->nb[1], - 0); - cb(kv_compressed, "kv_compressed", il); - - // and {n_embd_head_qk_rope, n_tokens} - struct ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens, - kv_pe_compresseed->nb[1], - kv_pe_compresseed->nb[1], - ggml_row_size(kv_pe_compresseed->type, kv_lora_rank)); - cb(k_pe, "k_pe", il); - - // TODO: the CUDA backend used to not support non-cont. (RMS) norm, investigate removing ggml_cont - kv_compressed = ggml_cont(ctx0, kv_compressed); - kv_compressed = llm_build_norm(ctx0, kv_compressed, hparams, - model.layers[il].attn_kv_a_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(kv_compressed, "kv_compressed", il); - - // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens} - struct ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed); - cb(kv, "kv", il); - - // split into {n_head * n_embd_head_qk_nope, n_tokens} - struct ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens, - ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v), - ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)), - 0); - cb(k_nope, "k_nope", il); - - // and {n_head * n_embd_head_v, n_tokens} - struct ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens, - ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)), - ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head), - ggml_row_size(kv->type, (n_embd_head_qk_nope))); - cb(v_states, "v_states", il); - - v_states = ggml_cont(ctx0, v_states); - cb(v_states, "v_states", il); - - v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens, - ggml_row_size(kv->type, hparams.n_embd_head_v * n_head), - 0); - cb(v_states, "v_states", il); - - q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this - q_pe = ggml_rope_ext( - ctx0, q_pe, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor_scaled, beta_fast, beta_slow - ); - cb(q_pe, "q_pe", il); - - // shared RoPE key - k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this - k_pe = ggml_rope_ext( - ctx0, k_pe, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor_scaled, beta_fast, beta_slow - ); - cb(k_pe, "k_pe", il); - - struct ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0); - cb(q_states, "q_states", il); - - struct ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0); - cb(k_states, "k_states", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - k_states, v_states, q_states, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - n_tokens = n_outputs; - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - if ((uint32_t) il < hparams.n_layer_dense_lead) { - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } else { - // MoE branch - ggml_tensor * moe_out = - llm_build_moe_ffn(ctx0, lctx, cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - model.layers[il].ffn_exp_probs_b, - n_expert, n_expert_used, - LLM_FFN_SILU, hparams.expert_weights_norm, - true, hparams.expert_weights_scale, - (enum llama_expert_gating_func_type) hparams.expert_gating_func, - cb, il); - cb(moe_out, "ffn_moe_out", il); - - // FFN shared expert - { - ggml_tensor * ffn_shexp = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up_shexp, NULL, NULL, - model.layers[il].ffn_gate_shexp, NULL, NULL, - model.layers[il].ffn_down_shexp, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(ffn_shexp, "ffn_shexp", il); - - cur = ggml_add(ctx0, moe_out, ffn_shexp); - cb(cur, "ffn_out", il); - } - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = ggml_mul_mat(ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_bitnet() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - if (model.layers[il].wq_scale) { - Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_scale); - } - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - // B1.K - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - if (model.layers[il].wk_scale) { - Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_scale); - } - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - // B1.V - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - if (model.layers[il].wv_scale) { - Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_scale); - } - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - NULL, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - - cur = llm_build_norm(ctx0, cur, hparams, - model.layers[il].attn_sub_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_sub_norm", il); - - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo, cur); - if (model.layers[il].wo_scale) { - cur = ggml_mul(ctx0, cur, model.layers[il].wo_scale); - } - if (model.layers[il].bo) { - cur = ggml_add(ctx0, cur, model.layers[il].bo); - } - cb(cur, "attn_o_out", il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward forward - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_scale, - model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_scale, - NULL, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_sub_out", il); - - cur = llm_build_norm(ctx0, cur, hparams, - model.layers[il].ffn_sub_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_sub_norm", il); - - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].ffn_down, cur); - if (model.layers[il].ffn_down_scale) { - cur = ggml_mul(ctx0, cur, model.layers[il].ffn_down_scale); - } - cb(cur, "ffn_down", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - // FIXME: do not use model.tok_embd directly, duplicate as model.output - cur = llm_build_lora_mm(lctx, ctx0, model.tok_embd, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - return gf; - } - - struct ggml_cgraph * build_t5_enc() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - // mutable variable, needed during the last layer of the computation to skip unused tokens - int32_t n_tokens = this->n_tokens; - - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - GGML_ASSERT(lctx.is_encoding); - struct ggml_tensor * pos_bucket_enc = llm_build_pos_bucket(false); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask_enc = build_inp_KQ_mask(false); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm_enc, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq_enc, cur); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk_enc, cur); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv_enc, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - - struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3)); - - struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); - cb(kq, "kq", il); - - struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc; - struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_enc, attn_rel_b); - struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias); - cb(kq_b, "kq_b", il); - - kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_enc, 1.0f, hparams.f_max_alibi_bias); - cb(kq, "kq_soft_max_ext", il); - - struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens))); - cb(v, "v", il); - - struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_tokens, n_embd_head, n_head_kv), kq); - cb(kqv, "kqv", il); - - struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); - cb(kqv_merged, "kqv_merged", il); - - cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); - cb(cur, "kqv_merged_cont", il); - - ggml_build_forward_expand(gf, cur); - - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo_enc, cur); - cb(cur, "kqv_out", il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - n_tokens = n_outputs; - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm_enc, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - // T5 uses relu, flan-T5 uses gelu-gated - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up_enc, NULL, NULL, - model.layers[il].ffn_gate_enc, NULL, NULL, - model.layers[il].ffn_down_enc, NULL, NULL, - NULL, - model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU, - model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ, - cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - ggml_tensor * layer_dir = lctx.cvec.tensor_for(il); - if (layer_dir != nullptr) { - cur = ggml_add(ctx0, cur, layer_dir); - } - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - cb(cur, "result_embd", -1); - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm_enc, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_t5_dec() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - // mutable variable, needed during the last layer of the computation to skip unused tokens - int32_t n_tokens = this->n_tokens; - - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - GGML_ASSERT(!lctx.is_encoding); - GGML_ASSERT(n_outputs_enc > 0 && "call llama_encode() first"); - - struct ggml_tensor * embd_enc = llm_build_inp_embd_enc(); - struct ggml_tensor * pos_bucket_dec = llm_build_pos_bucket(true); - - struct ggml_tensor * KQ_mask_dec = build_inp_KQ_mask(); - struct ggml_tensor * KQ_mask_cross = llm_build_inp_KQ_mask_cross(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - llm_build_kv_store(ctx0, hparams, cparams, kv_self, gf, Kcur, Vcur, n_tokens, kv_head, cb, il); - - struct ggml_tensor * k = - ggml_view_3d(ctx0, kv_self.k_l[il], - n_embd_head_k, n_kv, n_head_kv, - ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa), - ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k), - 0); - cb(k, "k", il); - - struct ggml_tensor * v = - ggml_view_3d(ctx0, kv_self.v_l[il], - n_kv, n_embd_head_v, n_head_kv, - ggml_element_size(kv_self.v_l[il])*n_ctx, - ggml_element_size(kv_self.v_l[il])*n_ctx*n_embd_head_v, - 0); - cb(v, "v", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - - struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - - struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); - cb(kq, "kq", il); - - struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b; - struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_dec, attn_rel_b); - struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias); - cb(kq_b, "kq_b", il); - - kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_dec, 1.0f, hparams.f_max_alibi_bias); - cb(kq, "kq_soft_max_ext", il); - - struct ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq); - cb(kqv, "kqv", il); - - struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); - cb(kqv_merged, "kqv_merged", il); - - cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); - cb(cur, "kqv_merged_cont", il); - - ggml_build_forward_expand(gf, cur); - - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo, cur); - cb(cur, "kqv_out", il); - } - - cur = ggml_add(ctx0, cur, inpSA); - cb(cur, "cross_inp", il); - - struct ggml_tensor * inpCA = cur; - - // norm - cur = llm_build_norm(ctx0, cur, hparams, - model.layers[il].attn_norm_cross, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm_cross", il); - - // cross-attention - { - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq_cross, cur); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk_cross, embd_enc); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv_cross, embd_enc); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc); - - struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3)); - - struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); - cb(kq, "kq", il); - - kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias); - cb(kq, "kq_soft_max_ext", il); - - struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc))); - cb(v, "v", il); - - struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq); - cb(kqv, "kqv", il); - - struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); - cb(kqv_merged, "kqv_merged", il); - - cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); - cb(cur, "kqv_merged_cont", il); - - ggml_build_forward_expand(gf, cur); - - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo_cross, cur); - cb(cur, "kqv_out", il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - n_tokens = n_outputs; - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - // T5 uses relu, flan-T5 uses gelu-gated - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU, - model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ, - cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - ggml_tensor * layer_dir = lctx.cvec.tensor_for(il); - if (layer_dir != nullptr) { - cur = ggml_add(ctx0, cur, layer_dir); - } - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - cb(cur, "result_embd", -1); - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_jais() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - - struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*cur->nb[0]*(n_embd))); - struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd))); - struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd + n_embd_gqa))); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/float(n_embd_head), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // add the input - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - // FF - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } - - inpL = ggml_add(ctx0, cur, ffn_inp); - cb(inpL, "l_out", il); - } - - cur = llm_build_norm(ctx0, inpL, hparams, - model.output_norm, - model.output_norm_b, - LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_chatglm() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - struct ggml_tensor * Qcur = nullptr; - struct ggml_tensor * Kcur = nullptr; - struct ggml_tensor * Vcur = nullptr; - if (model.layers[il].wqkv == nullptr) { - Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - } - Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - } - Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - } - } else { - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - if (model.layers[il].bqkv) { - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - } - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - } - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - //printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor); - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur_rope", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur_rope", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - // Add the input - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // FF - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, - NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SWIGLU, LLM_FFN_SEQ, cb, il); - cb(cur, "ffn_out", il); - - } - - inpL = ggml_add(ctx0, cur, ffn_inp); - cb(inpL, "l_out", il); - } - - cur = llm_build_norm(ctx0, inpL, hparams, - model.output_norm, - NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_nemotron() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - //GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_RELU_SQR, LLM_FFN_SEQ, cb, il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, model.output_norm_b, - LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_exaone() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - // mutable variable, needed during the last layer of the computation to skip unused tokens - int32_t n_tokens = this->n_tokens; - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // rope freq factors for llama3; may return nullptr for llama2 and other models - struct ggml_tensor * rope_factors = build_rope_factors(il); - - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - n_tokens = n_outputs; - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - ggml_cgraph * build_rwkv6() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - // Token shift state dimensions should be 2 * n_emb - GGML_ASSERT(n_embd == hparams.n_embd_k_s() / 2); - - const int64_t n_seqs = ubatch.n_seqs; - const int64_t n_seq_tokens = ubatch.n_seq_tokens; - const int64_t n_tokens = ubatch.n_tokens; - GGML_ASSERT(n_seqs != 0); - GGML_ASSERT(ubatch.equal_seqs); - GGML_ASSERT(n_tokens == n_seq_tokens * n_seqs); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - struct ggml_tensor * state_copy = build_inp_s_copy(); - struct ggml_tensor * state_mask = build_inp_s_mask(); - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - inpL = llm_build_norm(ctx0, inpL, hparams, model.tok_norm, model.tok_norm_b, LLM_NORM, cb, -1); - - for (int il = 0; il < n_layer; ++il) { - const llama_layer * layer = &model.layers[il]; - - // (ab)using the KV cache to store the states - struct ggml_tensor * token_shift = llm_build_copy_mask_state(ctx0, - gf, kv_self.k_l[il], state_copy, state_mask, - hparams.n_embd_k_s(), kv_self.size, kv_head, n_kv, n_seqs); - struct ggml_tensor * wkv_states = llm_build_copy_mask_state(ctx0, - gf, kv_self.v_l[il], state_copy, state_mask, - hparams.n_embd_v_s(), kv_self.size, kv_head, n_kv, n_seqs); - - cur = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); - token_shift = ggml_reshape_3d(ctx0, token_shift, n_embd, 2, n_seqs); - - struct ggml_tensor * att_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0); - struct ggml_tensor * ffn_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], n_embd * ggml_element_size(token_shift)); - - struct ggml_tensor * x_norm_att = llm_build_norm(ctx0, cur, hparams, layer->attn_norm, layer->attn_norm_b, LLM_NORM, cb, il); - struct ggml_tensor * x_prev = ggml_concat( - ctx0, - att_shift, - ggml_view_3d(ctx0, x_norm_att, n_embd, n_seq_tokens - 1, n_seqs, x_norm_att->nb[1], x_norm_att->nb[2], 0), - 1 - ); - - cur = ggml_add(ctx0, cur, llm_build_rwkv6_time_mix(lctx, ctx0, layer, x_norm_att, x_prev, &wkv_states, hparams.wkv_head_size, n_embd / hparams.wkv_head_size)); - ggml_build_forward_expand(gf, cur); - ggml_build_forward_expand( - gf, - ggml_cpy( - ctx0, - wkv_states, - ggml_view_1d( - ctx0, - kv_self.v_l[il], - hparams.n_embd_v_s() * n_seqs, - hparams.n_embd_v_s() * kv_head * ggml_element_size(kv_self.v_l[il]) - ) - ) - ); - - struct ggml_tensor * x_norm_ffn = llm_build_norm(ctx0, cur, hparams, layer->attn_norm_2, layer->attn_norm_2_b, LLM_NORM, cb, il); - x_prev = ggml_concat( - ctx0, - ffn_shift, - ggml_view_3d(ctx0, x_norm_ffn, n_embd, n_seq_tokens - 1, n_seqs, x_norm_ffn->nb[1], x_norm_ffn->nb[2], 0), - 1 - ); - cur = ggml_add(ctx0, cur, llm_build_rwkv6_channel_mix(lctx, ctx0, layer, x_norm_ffn, x_prev)); - ggml_build_forward_expand(gf, cur); - - struct ggml_tensor * last_norm_att = ggml_view_3d(ctx0, x_norm_att, n_embd, 1, n_seqs, x_norm_att->nb[1], x_norm_att->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(x_norm_att)); - struct ggml_tensor * last_norm_ffn = ggml_view_3d(ctx0, x_norm_ffn, n_embd, 1, n_seqs, x_norm_ffn->nb[1], x_norm_ffn->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(x_norm_ffn)); - - token_shift = ggml_concat(ctx0, last_norm_att, last_norm_ffn, 1); - - ggml_build_forward_expand( - gf, - ggml_cpy( - ctx0, - ggml_view_1d(ctx0, token_shift, n_embd * n_seqs * 2, 0), - ggml_view_1d(ctx0, kv_self.k_l[il], hparams.n_embd_k_s() * n_seqs, hparams.n_embd_k_s() * kv_head * ggml_element_size(kv_self.k_l[il])) - ) - ); - - if (hparams.rescale_every_n_layers != 0 && (il + 1) % hparams.rescale_every_n_layers == 0) { - cur = ggml_scale(ctx0, cur, 0.5F); - } - - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - - cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, model.output_norm_b, LLM_NORM, cb, -1); - cb(cur, "result_norm", -1); - - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - // ref: https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1/blob/main/modeling_rwkv6qwen2.py - ggml_cgraph * build_rwkv6qwen2() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - GGML_ASSERT(n_embd == hparams.n_embd_k_s()); - - const int64_t n_seqs = ubatch.n_seqs; - const int64_t n_seq_tokens = ubatch.n_seq_tokens; - const int64_t n_tokens = ubatch.n_tokens; - GGML_ASSERT(n_seqs != 0); - GGML_ASSERT(ubatch.equal_seqs); - GGML_ASSERT(n_tokens == n_seq_tokens * n_seqs); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - struct ggml_tensor * state_copy = build_inp_s_copy(); - struct ggml_tensor * state_mask = build_inp_s_mask(); - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - for (int il = 0; il < n_layer; ++il) { - const llama_layer * layer = &model.layers[il]; - - // (ab)using the KV cache to store the states - struct ggml_tensor * token_shift = llm_build_copy_mask_state(ctx0, - gf, kv_self.k_l[il], state_copy, state_mask, - hparams.n_embd_k_s(), kv_self.size, kv_head, n_kv, n_seqs); - struct ggml_tensor * wkv_states = llm_build_copy_mask_state(ctx0, - gf, kv_self.v_l[il], state_copy, state_mask, - hparams.n_embd_v_s(), kv_self.size, kv_head, n_kv, n_seqs); - - cur = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); - token_shift = ggml_reshape_3d(ctx0, token_shift, n_embd, 1, n_seqs); - - struct ggml_tensor * x_norm_att = llm_build_norm(ctx0, cur, hparams, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, cb, il); - struct ggml_tensor * x_prev = ggml_concat( - ctx0, - token_shift, - ggml_view_3d(ctx0, x_norm_att, n_embd, n_seq_tokens - 1, n_seqs, x_norm_att->nb[1], x_norm_att->nb[2], 0), - 1 - ); - - struct ggml_tensor * last_norm_att = ggml_view_3d(ctx0, x_norm_att, n_embd, 1, n_seqs, x_norm_att->nb[1], x_norm_att->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(x_norm_att)); - ggml_build_forward_expand( - gf, - ggml_cpy( - ctx0, - ggml_view_1d(ctx0, last_norm_att, n_embd * n_seqs, 0), - ggml_view_1d(ctx0, kv_self.k_l[il], hparams.n_embd_k_s() * n_seqs, hparams.n_embd_k_s() * kv_head * ggml_element_size(kv_self.k_l[il])) - ) - ); - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, llm_build_rwkv6_time_mix(lctx, ctx0, layer, x_norm_att, x_prev, &wkv_states, hparams.wkv_head_size, hparams.n_head_kv())); - ggml_build_forward_expand(gf, ffn_inp); - ggml_build_forward_expand( - gf, - ggml_cpy( - ctx0, - wkv_states, - ggml_view_1d( - ctx0, - kv_self.v_l[il], - hparams.n_embd_v_s() * n_seqs, - hparams.n_embd_v_s() * kv_head * ggml_element_size(kv_self.v_l[il]) - ) - ) - ); - - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - - cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, model.output_norm_b, LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - // ref: https://github.com/facebookresearch/chameleon - // based on the original build_llama() function, changes: - // * qk-norm - // * swin-norm - // * removed bias - // * removed MoE - struct ggml_cgraph * build_chameleon() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - // mutable variable, needed during the last layer of the computation to skip unused tokens - int32_t n_tokens = this->n_tokens; - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - if (hparams.swin_norm) { - cur = inpL; - } else { - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - } - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - if (model.layers[il].attn_q_norm) { - Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens, - ggml_element_size(Qcur) * n_embd_head, - ggml_element_size(Qcur) * n_embd_head * n_head, - 0); - cb(Qcur, "Qcur", il); - - Qcur = llm_build_norm(ctx0, Qcur, hparams, - model.layers[il].attn_q_norm, - model.layers[il].attn_q_norm_b, - LLM_NORM, cb, il); - cb(Qcur, "Qcur", il); - } - - if (model.layers[il].attn_k_norm) { - Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens, - ggml_element_size(Kcur) * n_embd_head, - ggml_element_size(Kcur) * n_embd_head * n_head_kv, - 0); - cb(Kcur, "Kcur", il); - - Kcur = llm_build_norm(ctx0, Kcur, hparams, - model.layers[il].attn_k_norm, - model.layers[il].attn_k_norm_b, - LLM_NORM, cb, il); - cb(Kcur, "Kcur", il); - } - - Qcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, nullptr, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - - if (hparams.swin_norm) { - cur = llm_build_norm(ctx0, cur, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - } - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - n_tokens = n_outputs; - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - if (!hparams.swin_norm) { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - } - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - - if (hparams.swin_norm) { - cur = llm_build_norm(ctx0, cur, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output_with_img_logits", -1); - - // TODO: this suppresses the output of image tokens, which is required to enable text-only outputs. - // Needs to be removed once image outputs are supported. - int img_token_end_idx = 8196; - int img_token_start_idx = 4; - int num_img_tokens = img_token_end_idx - img_token_start_idx; - // creates 1d tensor of size num_img_tokens and values -FLT_MAX, - // which ensures that text token values are always at least larger than image token values - struct ggml_tensor * img_logits = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, num_img_tokens); - img_logits = ggml_clamp(ctx0, img_logits, -FLT_MAX, -FLT_MAX); - cb(img_logits, "img_logits", -1); - cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - - struct ggml_cgraph * build_wavtokenizer_dec() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - cur = ggml_cont(ctx0, ggml_transpose(ctx0, inpL)); - - cur = ggml_conv_1d_ph(ctx0, model.conv1d, cur, 1, 1); - cur = ggml_add(ctx0, cur, model.conv1d_b); - - // posnet - for (uint32_t il = 0; il < hparams.posnet.n_layer; ++il) { - const auto & layer = model.layers[il].posnet; - - inpL = cur; - - switch (il) { - case 0: - case 1: - case 3: - case 4: - { - cur = llm_build_norm(ctx0, cur, hparams, - layer.norm1, - layer.norm1_b, - LLM_NORM_GROUP, cb, 0); - - cur = ggml_mul(ctx0, ggml_sigmoid(ctx0, cur), cur); - - cur = ggml_conv_1d_ph(ctx0, layer.conv1, cur, 1, 1); - cur = ggml_add(ctx0, cur, layer.conv1_b); - - cur = llm_build_norm(ctx0, cur, hparams, - layer.norm2, - layer.norm2_b, - LLM_NORM_GROUP, cb, 0); - - cur = ggml_mul(ctx0, ggml_sigmoid(ctx0, cur), cur); - - cur = ggml_conv_1d_ph(ctx0, layer.conv2, cur, 1, 1); - cur = ggml_add(ctx0, cur, layer.conv2_b); - - cur = ggml_add(ctx0, cur, inpL); - } break; - case 2: - { - cur = llm_build_norm(ctx0, cur, hparams, - layer.attn_norm, - layer.attn_norm_b, - LLM_NORM_GROUP, cb, 0); - - struct ggml_tensor * q; - struct ggml_tensor * k; - struct ggml_tensor * v; - - q = ggml_conv_1d_ph(ctx0, layer.attn_q, cur, 1, 1); - k = ggml_conv_1d_ph(ctx0, layer.attn_k, cur, 1, 1); - v = ggml_conv_1d_ph(ctx0, layer.attn_v, cur, 1, 1); - - q = ggml_add(ctx0, q, layer.attn_q_b); - k = ggml_add(ctx0, k, layer.attn_k_b); - v = ggml_add(ctx0, v, layer.attn_v_b); - - q = ggml_cont(ctx0, ggml_transpose(ctx0, q)); - k = ggml_cont(ctx0, ggml_transpose(ctx0, k)); - - struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); - - kq = ggml_soft_max_ext(ctx0, kq, nullptr, 1.0f/sqrtf(float(hparams.posnet.n_embd)), 0.0f); - - cur = ggml_mul_mat(ctx0, kq, v); - - cur = ggml_conv_1d_ph(ctx0, layer.attn_o, cur, 1, 1); - cur = ggml_add(ctx0, cur, layer.attn_o_b); - - cur = ggml_add(ctx0, cur, inpL); - } break; - case 5: - { - cur = llm_build_norm(ctx0, cur, hparams, - layer.norm, - layer.norm_b, - LLM_NORM_GROUP, cb, 0); - } break; - default: GGML_ABORT("unknown posnet layer"); - }; - } - - cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); - - cur = llm_build_norm(ctx0, cur, hparams, - model.tok_norm, - model.tok_norm_b, - LLM_NORM, cb, -1); - - cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); - - inpL = cur; - - // convnext - for (uint32_t il = 0; il < hparams.convnext.n_layer; ++il) { - const auto & layer = model.layers[il].convnext; - - cur = inpL; - - cur = ggml_conv_1d_dw_ph(ctx0, layer.dw, cur, 1, 1); - cur = ggml_add(ctx0, cur, layer.dw_b); - - cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); - - cur = llm_build_norm(ctx0, cur, hparams, - layer.norm, - layer.norm_b, - LLM_NORM, cb, -1); - - cur = llm_build_ffn(ctx0, lctx, cur, - layer.pw1, layer.pw1_b, NULL, - NULL, NULL, NULL, - layer.pw2, layer.pw2_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); - - cur = ggml_mul(ctx0, cur, layer.gamma); - - cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); - - inpL = ggml_add(ctx0, cur, inpL); - } - - cur = inpL; - - cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, - model.output_norm_b, - LLM_NORM, cb, -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - - cur = ggml_add(ctx0, cur, model.output_b); - cb(cur, "result_embd", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } -}; - -static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector & ids) { - llama_ubatch dummy = {}; - dummy.equal_seqs = true; - - llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { }; - - struct llm_build_context llm(lctx, dummy, cb, false); - - llm.init(); - - struct ggml_cgraph * result = llm.build_defrag(ids); - - llm.free(); - - return result; -} - -static struct ggml_cgraph * llama_build_graph_k_shift(llama_context & lctx) { - llama_ubatch dummy = {}; - dummy.equal_seqs = true; - - llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { }; - - struct llm_build_context llm(lctx, dummy, cb, false); - - llm.init(); - - struct ggml_cgraph * result = llm.build_k_shift(); - - llm.free(); - - return result; -} - -static struct ggml_cgraph * llama_build_graph( - llama_context & lctx, - const llama_ubatch & ubatch, - bool worst_case) { - const auto & model = lctx.model; - - // this callback allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.) - llm_build_cb cb = [&](struct ggml_tensor * cur, const char * name, int il) { - if (il >= 0) { - ggml_format_name(cur, "%s-%d", name, il); - } else { - ggml_set_name(cur, name); - } - - if (!lctx.cparams.offload_kqv) { - if (strcmp(name, "kqv_merged_cont") == 0) { - // all nodes between the KV store and the attention output are run on the CPU - ggml_backend_sched_set_tensor_backend(lctx.sched.get(), cur, lctx.backend_cpu); - } - } - - // norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends - // FIXME: fix in ggml_backend_sched - const bool full_offload = lctx.model.params.n_gpu_layers > (int) lctx.model.hparams.n_layer; - if (ubatch.n_tokens < 32 || full_offload) { - if (il != -1 && strcmp(name, "norm") == 0) { - const auto & dev_layer = lctx.model.dev_layer(il); - for (auto & backend : lctx.backends) { - if (ggml_backend_get_device(backend.get()) == dev_layer) { - if (ggml_backend_supports_op(backend.get(), cur)) { - ggml_backend_sched_set_tensor_backend(lctx.sched.get(), cur, backend.get()); - } - } - } - } - } - }; - - struct ggml_cgraph * result = NULL; - - struct llm_build_context llm(lctx, ubatch, cb, worst_case); - - llm.init(); - - switch (model.arch) { - case LLM_ARCH_LLAMA: - case LLM_ARCH_MINICPM: - case LLM_ARCH_GRANITE: - case LLM_ARCH_GRANITE_MOE: - { - result = llm.build_llama(); - } break; - case LLM_ARCH_DECI: - { - result = llm.build_deci(); - } break; - case LLM_ARCH_BAICHUAN: - { - result = llm.build_baichuan(); - } break; - case LLM_ARCH_FALCON: - { - result = llm.build_falcon(); - } break; - case LLM_ARCH_GROK: - { - result = llm.build_grok(); - } break; - case LLM_ARCH_STARCODER: - { - result = llm.build_starcoder(); - } break; - case LLM_ARCH_REFACT: - { - result = llm.build_refact(); - } break; - case LLM_ARCH_BERT: - case LLM_ARCH_JINA_BERT_V2: - case LLM_ARCH_NOMIC_BERT: - { - result = llm.build_bert(); - } break; - case LLM_ARCH_BLOOM: - { - result = llm.build_bloom(); - } break; - case LLM_ARCH_MPT: - { - result = llm.build_mpt(); - } break; - case LLM_ARCH_STABLELM: - { - result = llm.build_stablelm(); - } break; - case LLM_ARCH_QWEN: - { - result = llm.build_qwen(); - } break; - case LLM_ARCH_QWEN2: - { - result = llm.build_qwen2(); - } break; - case LLM_ARCH_QWEN2VL: - { - lctx.n_pos_per_token = 4; - result = llm.build_qwen2vl(); - } break; - case LLM_ARCH_QWEN2MOE: - { - result = llm.build_qwen2moe(); - } break; - case LLM_ARCH_PHI2: - { - result = llm.build_phi2(); - } break; - case LLM_ARCH_PHI3: - case LLM_ARCH_PHIMOE: - { - result = llm.build_phi3(); - } break; - case LLM_ARCH_PLAMO: - { - result = llm.build_plamo(); - } break; - case LLM_ARCH_GPT2: - { - result = llm.build_gpt2(); - } break; - case LLM_ARCH_CODESHELL: - { - result = llm.build_codeshell(); - } break; - case LLM_ARCH_ORION: - { - result = llm.build_orion(); - } break; - case LLM_ARCH_INTERNLM2: - { - result = llm.build_internlm2(); - } break; - case LLM_ARCH_MINICPM3: - { - result = llm.build_minicpm3(); - } break; - case LLM_ARCH_GEMMA: - { - result = llm.build_gemma(); - } break; - case LLM_ARCH_GEMMA2: - { - result = llm.build_gemma2(); - } break; - case LLM_ARCH_STARCODER2: - { - result = llm.build_starcoder2(); - } break; - case LLM_ARCH_MAMBA: - { - result = llm.build_mamba(); - } break; - case LLM_ARCH_XVERSE: - { - result = llm.build_xverse(); - } break; - case LLM_ARCH_COMMAND_R: - { - result = llm.build_command_r(); - } break; - case LLM_ARCH_COHERE2: - { - result = llm.build_cohere2(); - } break; - case LLM_ARCH_DBRX: - { - result = llm.build_dbrx(); - } break; - case LLM_ARCH_OLMO: - { - result = llm.build_olmo(); - } break; - case LLM_ARCH_OLMO2: - { - result = llm.build_olmo2(); - } break; - case LLM_ARCH_OLMOE: - { - result = llm.build_olmoe(); - } break; - case LLM_ARCH_OPENELM: - { - result = llm.build_openelm(); - } break; - case LLM_ARCH_GPTNEOX: - { - result = llm.build_gptneox(); - } break; - case LLM_ARCH_ARCTIC: - { - result = llm.build_arctic(); - } break; - case LLM_ARCH_DEEPSEEK: - { - result = llm.build_deepseek(); - } break; - case LLM_ARCH_DEEPSEEK2: - { - result = llm.build_deepseek2(); - } break; - case LLM_ARCH_CHATGLM: - { - result = llm.build_chatglm(); - } break; - case LLM_ARCH_BITNET: - { - result = llm.build_bitnet(); - } break; - case LLM_ARCH_T5: - { - if (lctx.is_encoding) { - result = llm.build_t5_enc(); - } else { - result = llm.build_t5_dec(); - } - } break; - case LLM_ARCH_T5ENCODER: - { - result = llm.build_t5_enc(); - } break; - case LLM_ARCH_JAIS: - { - result = llm.build_jais(); - } break; - case LLM_ARCH_NEMOTRON: - { - result = llm.build_nemotron(); - } break; - case LLM_ARCH_EXAONE: - { - result = llm.build_exaone(); - } break; - case LLM_ARCH_RWKV6: - { - result = llm.build_rwkv6(); - } break; - case LLM_ARCH_RWKV6QWEN2: - { - result = llm.build_rwkv6qwen2(); - } break; - case LLM_ARCH_CHAMELEON: - { - result = llm.build_chameleon(); - } break; - case LLM_ARCH_WAVTOKENIZER_DEC: - { - result = llm.build_wavtokenizer_dec(); - } break; - default: - GGML_ABORT("fatal error"); - } - - // add on pooling layer - if (lctx.cparams.embeddings) { - result = llm.append_pooling(result); - } - - llm.free(); - - return result; -} - -// returns the result of ggml_backend_sched_graph_compute_async execution -static enum ggml_status llama_graph_compute( - llama_context & lctx, - ggml_cgraph * gf, - int n_threads, - ggml_threadpool * threadpool) { - if (lctx.backend_cpu != nullptr) { - auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(lctx.backend_cpu)); - auto * set_threadpool_fn = (decltype(ggml_backend_cpu_set_threadpool) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool"); - set_threadpool_fn(lctx.backend_cpu, threadpool); - } - - // set the number of threads for all the backends - for (const auto & set_n_threads_fn : lctx.set_n_threads_fns) { - set_n_threads_fn.second(set_n_threads_fn.first, n_threads); - } - - auto status = ggml_backend_sched_graph_compute_async(lctx.sched.get(), gf); - if (status != GGML_STATUS_SUCCESS) { - LLAMA_LOG_ERROR("%s: ggml_backend_sched_graph_compute_async failed with error %d\n", __func__, status); - } - - // fprintf(stderr, "splits: %d\n", ggml_backend_sched_get_n_splits(lctx.sched)); - - return status; -} - -static int llama_prepare_sbatch( - llama_context & lctx, - const llama_batch & batch, - uint32_t & n_outputs) { - const auto & model = lctx.model; - const auto & hparams = model.hparams; - const auto & cparams = lctx.cparams; - - const uint32_t n_tokens_all = batch.n_tokens; - const int64_t n_embd = hparams.n_embd; - - // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens - const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE; - - GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT - if (batch.token) { - for (uint32_t i = 0; i < n_tokens_all; ++i) { - if (batch.token[i] < 0 || uint32_t(batch.token[i]) >= model.vocab.n_tokens()) { - LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]); - return -1; - } - } - } - GGML_ASSERT(n_tokens_all <= cparams.n_batch); - GGML_ASSERT((cparams.causal_attn || cparams.n_ubatch >= n_tokens_all) && "non-causal attention requires n_ubatch >= n_tokens"); - - lctx.n_queued_tokens += n_tokens_all; - lctx.embd_seq.clear(); - - // count outputs - if (batch.logits && !embd_pooled) { - for (uint32_t i = 0; i < n_tokens_all; ++i) { - n_outputs += batch.logits[i] != 0; - } - } else if (lctx.logits_all || embd_pooled) { - n_outputs = n_tokens_all; - } else { - // keep last output only - n_outputs = 1; - } - - lctx.sbatch.from_batch(batch, n_embd, - /* simple_split */ !lctx.kv_self.recurrent, - /* logits_all */ n_outputs == n_tokens_all); - - // reserve output buffer - if (llama_output_reserve(lctx, n_outputs) < n_outputs) { - LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_outputs); - return -2; - }; - - return 0; -} - -static int llama_prepare_ubatch( - llama_context & lctx, - llama_kv_slot_restorer & kv_slot_restorer, - llama_ubatch & ubatch, - const uint32_t n_outputs, - const uint32_t n_tokens_all) { - GGML_ASSERT(lctx.sbatch.n_tokens > 0); - - auto & kv_self = lctx.kv_self; - const auto & cparams = lctx.cparams; - const auto & hparams = lctx.model.hparams; - - // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens - const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE; - - if (lctx.kv_self.recurrent) { - if (embd_pooled) { - // Pooled embeddings cannot be split across ubatches (yet) - ubatch = lctx.sbatch.split_seq(cparams.n_ubatch); - } else { - // recurrent model architectures are easier to implement - // with equal-length sequences - ubatch = lctx.sbatch.split_equal(cparams.n_ubatch); - } - } else { - ubatch = lctx.sbatch.split_simple(cparams.n_ubatch); - } - - // count the outputs in this u_batch - { - int32_t n_outputs_new = 0; - - if (n_outputs == n_tokens_all) { - n_outputs_new = ubatch.n_tokens; - } else { - GGML_ASSERT(ubatch.output); - for (uint32_t i = 0; i < ubatch.n_tokens; i++) { - n_outputs_new += int32_t(ubatch.output[i] != 0); - } - } - - // needs to happen before the graph is built - lctx.n_outputs = n_outputs_new; - } - - // non-causal masks do not use the KV cache - if (hparams.causal_attn) { - llama_kv_cache_update(&lctx); - - // if we have enough unused cells before the current head -> - // better to start searching from the beginning of the cache, hoping to fill it - if (kv_self.head > kv_self.used + 2*ubatch.n_tokens) { - kv_self.head = 0; - } - - const auto slot = llama_kv_cache_find_slot(kv_self, ubatch); - if (!slot) { - return 1; - } - kv_slot_restorer.save(slot); - - if (!kv_self.recurrent) { - // a heuristic, to avoid attending the full cache if it is not yet utilized - // after enough generations, the benefit from this heuristic disappears - // if we start defragmenting the cache, the benefit from this will be more important - const uint32_t pad = llama_kv_cache_get_padding(cparams); - kv_self.n = std::min(kv_self.size, std::max(pad, GGML_PAD(llama_kv_cache_cell_max(kv_self), pad))); - //kv_self.n = llama_kv_cache_cell_max(kv_self); - } - } - - return 0; -} - -// decode a batch of tokens by evaluating the transformer -// in case of unsuccessful decoding (error or warning), -// the kv_cache state will be returned to its original state -// (for non-recurrent models) or cleaned (for recurrent models) -// -// - lctx: llama context -// - inp_batch: batch to evaluate -// -// return 0 on success -// return positive int on warning -// return negative int on error -// -static int llama_decode_impl( - llama_context & lctx, - llama_batch inp_batch) { - - lctx.is_encoding = false; - - if (inp_batch.n_tokens == 0) { - LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__); - return -1; - } - - // temporarily allocate memory for the input batch if needed - llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : lctx.kv_self.max_pos() + 1); - const llama_batch & batch = batch_allocr.batch; - - const auto & model = lctx.model; - const auto & vocab = model.vocab; - const auto & hparams = model.hparams; - const auto & cparams = lctx.cparams; - - if (lctx.t_compute_start_us == 0) { - lctx.t_compute_start_us = ggml_time_us(); - } - auto & kv_self = lctx.kv_self; - llama_kv_slot_restorer kv_slot_restorer(kv_self); - - const int64_t n_embd = hparams.n_embd; - const int64_t n_vocab = vocab.n_tokens(); - - uint32_t n_outputs = 0; - uint32_t n_outputs_prev = 0; - - { - const int ret = llama_prepare_sbatch(lctx, batch, n_outputs); - if (ret != 0) { - return ret; - } - } - - while (lctx.sbatch.n_tokens > 0) { - llama_ubatch ubatch; - { - const int ret = llama_prepare_ubatch(lctx, kv_slot_restorer, ubatch, n_outputs, batch.n_tokens); - if (ret != 0) { - return ret; - } - } - - const int n_threads = ubatch.n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch; - ggml_threadpool_t threadpool = ubatch.n_tokens == 1 ? lctx.threadpool : lctx.threadpool_batch; - - GGML_ASSERT(n_threads > 0); - - //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head); - - ggml_backend_sched_reset(lctx.sched.get()); - ggml_backend_sched_set_eval_callback(lctx.sched.get(), lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data); - - ggml_cgraph * gf = llama_build_graph(lctx, ubatch, false); - - // the output is always the last tensor in the graph - struct ggml_tensor * res = ggml_graph_node(gf, -1); - struct ggml_tensor * embd = ggml_graph_node(gf, -2); - - if (lctx.n_outputs == 0) { - // no output - res = nullptr; - embd = nullptr; - } else if (cparams.embeddings) { - res = nullptr; // do not extract logits for embedding case - embd = nullptr; - for (int i = ggml_graph_n_nodes(gf) - 1; i >= 0; --i) { - if (strcmp(ggml_graph_node(gf, i)->name, "result_embd_pooled") == 0) { - embd = ggml_graph_node(gf, i); - break; - } - } - GGML_ASSERT(embd != nullptr && "missing embeddings tensor"); - } else { - embd = nullptr; // do not extract embeddings when not needed - GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor"); - } - - // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs); - - ggml_backend_sched_alloc_graph(lctx.sched.get(), gf); - - llama_set_inputs(lctx, ubatch); - - const auto compute_status = llama_graph_compute(lctx, gf, n_threads, threadpool); - if (compute_status != GGML_STATUS_SUCCESS) { - kv_slot_restorer.restore(kv_self); - switch (compute_status) { - case GGML_STATUS_ABORTED: - return 2; - case GGML_STATUS_ALLOC_FAILED: - return -2; - case GGML_STATUS_FAILED: - default: - return -3; - } - } - - // update the kv ring buffer - { - kv_self.head += ubatch.n_tokens; - - // Ensure kv cache head points to a valid index. - if (kv_self.head >= kv_self.size) { - kv_self.head = 0; - } - } - - // plot the computation graph in dot format (for debugging purposes) - //if (n_past%100 == 0) { - // ggml_graph_dump_dot(gf, NULL, "llama.dot"); - //} - - // extract logits - if (res) { - ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(lctx.sched.get(), res); - GGML_ASSERT(backend_res != nullptr); - GGML_ASSERT(lctx.logits != nullptr); - - float * logits_out = lctx.logits + n_outputs_prev*n_vocab; - const int32_t n_outputs_new = lctx.n_outputs; - - if (n_outputs_new) { - GGML_ASSERT( n_outputs_prev + n_outputs_new <= n_outputs); - GGML_ASSERT((n_outputs_prev + n_outputs_new)*n_vocab <= (int64_t) lctx.logits_size); - ggml_backend_tensor_get_async(backend_res, res, logits_out, 0, n_outputs_new*n_vocab*sizeof(float)); - } - } - - // extract embeddings - if (embd) { - ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(lctx.sched.get(), embd); - GGML_ASSERT(backend_embd != nullptr); - - switch (cparams.pooling_type) { - case LLAMA_POOLING_TYPE_NONE: - { - // extract token embeddings - GGML_ASSERT(lctx.embd != nullptr); - float * embd_out = lctx.embd + n_outputs_prev*n_embd; - const int32_t n_outputs_new = lctx.n_outputs; - - if (n_outputs_new) { - GGML_ASSERT( n_outputs_prev + n_outputs_new <= n_outputs); - GGML_ASSERT((n_outputs_prev + n_outputs_new)*n_embd <= (int64_t) lctx.embd_size); - ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_outputs_new*n_embd*sizeof(float)); - } - } break; - case LLAMA_POOLING_TYPE_MEAN: - case LLAMA_POOLING_TYPE_CLS: - case LLAMA_POOLING_TYPE_LAST: - { - // extract sequence embeddings (cleared before processing each batch) - auto & embd_seq_out = lctx.embd_seq; - - for (uint32_t s = 0; s < ubatch.n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; - if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { - continue; - } - embd_seq_out[seq_id].resize(n_embd); - ggml_backend_tensor_get_async(backend_embd, embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float)); - } - } break; - case LLAMA_POOLING_TYPE_RANK: - { - // extract the rerank score - a single float per sequence - auto & embd_seq_out = lctx.embd_seq; - - for (uint32_t s = 0; s < ubatch.n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; - if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { - continue; - } - embd_seq_out[seq_id].resize(1); - ggml_backend_tensor_get_async(backend_embd, embd, embd_seq_out[seq_id].data(), (seq_id)*sizeof(float), sizeof(float)); - } - } break; - case LLAMA_POOLING_TYPE_UNSPECIFIED: - { - GGML_ABORT("unknown pooling type"); - } - } - } - n_outputs_prev += lctx.n_outputs; - } - - // set output mappings - { - bool sorted_output = true; - - GGML_ASSERT(lctx.sbatch.out_ids.size() == n_outputs); - - for (size_t i = 0; i < n_outputs; ++i) { - size_t out_id = lctx.sbatch.out_ids[i]; - lctx.output_ids[out_id] = i; - if (out_id != i) { - sorted_output = false; - } - } - - if (sorted_output) { - lctx.sbatch.out_ids.clear(); - } - } - - // set to total number of outputs in the batch, for use in llama_get_logits_ith - lctx.n_outputs = n_outputs; - - // wait for the computation to finish (automatically done when obtaining the model output) - //llama_synchronize(&lctx); - - // decide if we need to defrag the kv cache - if (cparams.causal_attn && cparams.defrag_thold > 0.0f) { - // - do not defrag small contexts (i.e. < 2048 tokens) - // - count the padding towards the number of used tokens - const float fragmentation = kv_self.n >= 2048 ? std::max(0.0f, 1.0f - float(kv_self.used + llama_kv_cache_get_padding(cparams))/float(kv_self.n)) : 0.0f; - - // queue defragmentation for next llama_kv_cache_update - if (fragmentation > cparams.defrag_thold) { - LLAMA_LOG_DEBUG("%s: fragmentation: %.2f - requesting defrag\n", __func__, fragmentation); - - llama_kv_cache_defrag(kv_self); - } - } - - // Reset state for the next token before backend sync, to allow the CPU activities in the reset to - // overlap with device computation. - ggml_backend_sched_reset(lctx.sched.get()); - - return 0; -} - -// encode a batch of tokens by evaluating the encoder part of the transformer -// -// - lctx: llama context -// - batch: batch to evaluate -// -// return 0 on success -// return positive int on warning -// return negative int on error -// -static int llama_encode_impl( - llama_context & lctx, - llama_batch inp_batch) { - - lctx.is_encoding = true; - - if (inp_batch.n_tokens == 0) { - LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__); - return -1; - } - - // temporary allocate memory for the input batch if needed - llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : lctx.kv_self.max_pos() + 1); - - const llama_batch & batch = batch_allocr.batch; - const uint32_t n_tokens = batch.n_tokens; - - const auto & model = lctx.model; - const auto & hparams = model.hparams; - const auto & cparams = lctx.cparams; - - GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT - - if (batch.token) { - for (uint32_t i = 0; i < n_tokens; ++i) { - if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= model.vocab.n_tokens()) { - LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]); - return -1; - } - } - } - - // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot - GGML_ASSERT(cparams.n_ubatch >= n_tokens && "encoder requires n_ubatch >= n_tokens"); - - if (lctx.t_compute_start_us == 0) { - lctx.t_compute_start_us = ggml_time_us(); - } - - lctx.n_queued_tokens += n_tokens; - - const int64_t n_embd = hparams.n_embd; - - lctx.sbatch.from_batch(batch, n_embd, /* simple_split */ true, /* logits_all */ true); - - const llama_ubatch ubatch = lctx.sbatch.split_simple(n_tokens); - - // reserve output buffer - if (llama_output_reserve(lctx, n_tokens) < n_tokens) { - LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_tokens); - return -2; - }; - - for (uint32_t i = 0; i < n_tokens; ++i) { - lctx.output_ids[i] = i; - } - - lctx.inp_embd_enc = NULL; - lctx.n_outputs = n_tokens; - - int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch; - ggml_threadpool_t threadpool = n_tokens == 1 ? lctx.threadpool : lctx.threadpool_batch; - - GGML_ASSERT(n_threads > 0); - - ggml_backend_sched_reset(lctx.sched.get()); - ggml_backend_sched_set_eval_callback(lctx.sched.get(), lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data); - - ggml_cgraph * gf = llama_build_graph(lctx, ubatch, false); - - // the output embeddings after the final encoder normalization - struct ggml_tensor * embd = nullptr; - - // there are two cases here - if (llama_model_has_decoder(&lctx.model)) { - // first case is an encoder-decoder T5 model where embeddings are passed to decoder - embd = ggml_graph_node(gf, -1); - GGML_ASSERT(strcmp(embd->name, "result_norm") == 0 && "missing result_output tensor"); - } else { - // second case is an encoder-only T5 model - if (cparams.embeddings) { - // only output embeddings if required - embd = ggml_graph_node(gf, -1); - if (strcmp(embd->name, "result_embd_pooled") != 0) { - embd = ggml_graph_node(gf, -2); - } - GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0 && "missing embeddings tensor"); - } - } - - ggml_backend_sched_alloc_graph(lctx.sched.get(), gf); - - llama_set_inputs(lctx, ubatch); - - const auto compute_status = llama_graph_compute(lctx, gf, n_threads, threadpool); - switch (compute_status) { - case GGML_STATUS_SUCCESS: - break; - case GGML_STATUS_ABORTED: - return 2; - case GGML_STATUS_ALLOC_FAILED: - return -2; - case GGML_STATUS_FAILED: - default: - return -3; - } - - // extract embeddings - if (embd) { - ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(lctx.sched.get(), embd); - GGML_ASSERT(backend_embd != nullptr); - - if (llama_model_has_decoder(&lctx.model)) { - lctx.embd_enc.resize(n_tokens*n_embd); - float * embd_out = lctx.embd_enc.data(); - - ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_tokens*n_embd*sizeof(float)); - GGML_ASSERT(!ubatch.equal_seqs); // TODO: handle equal splits - - // remember the sequence ids used during the encoding - needed for cross attention later - lctx.seq_ids_enc.resize(n_tokens); - for (uint32_t i = 0; i < n_tokens; i++) { - for (int s = 0; s < ubatch.n_seq_id[i]; s++) { - llama_seq_id seq_id = ubatch.seq_id[i][s]; - lctx.seq_ids_enc[i].insert(seq_id); - } - } - } else { - GGML_ASSERT(lctx.embd != nullptr); - - switch (cparams.pooling_type) { - case LLAMA_POOLING_TYPE_NONE: - { - // extract token embeddings - GGML_ASSERT(lctx.embd != nullptr); - float * embd_out = lctx.embd; - - GGML_ASSERT(n_tokens*n_embd <= (int64_t) lctx.embd_size); - ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_tokens*n_embd*sizeof(float)); - } break; - case LLAMA_POOLING_TYPE_MEAN: - case LLAMA_POOLING_TYPE_CLS: - case LLAMA_POOLING_TYPE_LAST: - { - // extract sequence embeddings - auto & embd_seq_out = lctx.embd_seq; - embd_seq_out.clear(); - - GGML_ASSERT(!ubatch.equal_seqs); // TODO: handle equal splits - - for (uint32_t i = 0; i < n_tokens; i++) { - const llama_seq_id seq_id = ubatch.seq_id[i][0]; - if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { - continue; - } - embd_seq_out[seq_id].resize(n_embd); - ggml_backend_tensor_get_async(backend_embd, embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float)); - } - } break; - case LLAMA_POOLING_TYPE_RANK: - { - // TODO: this likely should be the same logic as in llama_decoder_internal, but better to - // wait for an encoder model that requires this pooling type in order to test it - // https://github.com/ggerganov/llama.cpp/pull/9510 - GGML_ABORT("RANK pooling not implemented yet"); - } - case LLAMA_POOLING_TYPE_UNSPECIFIED: - { - GGML_ABORT("unknown pooling type"); - } - } - } - } - - // Reset state for the next token before backend sync, to allow the CPU activities in the reset to - // overlap with device computation. - ggml_backend_sched_reset(lctx.sched.get()); - - return 0; -} - -// find holes from the beginning of the KV cache and fill them by moving data from the end of the cache -static void llama_kv_cache_defrag_impl(struct llama_context & lctx) { - auto & kv_self = lctx.kv_self; - - const auto & hparams = lctx.model.hparams; - - const uint32_t n_layer = hparams.n_layer; - - const uint32_t n_kv = llama_kv_cache_cell_max(kv_self); - const uint32_t n_used = kv_self.used; - - assert(n_used <= n_kv); - - //const int64_t t_start = ggml_time_us(); - - // number of cells moved - uint32_t n_moves = 0; - - // each move requires 6*n_layer tensors (see build_defrag) - // - source view, destination view, copy operation - // - x2 for keys and values - //const uint32_t max_moves = model.max_nodes()/(6*n_layer); - // TODO: tmp fix https://github.com/ggerganov/llama.cpp/issues/6685#issuecomment-2057579516 - const uint32_t max_moves = (lctx.model.max_nodes() - 2*n_layer)/(6*n_layer); - - // determine which KV cells to move where - // - // cell i moves to ids[i] - // - // if ids[i] == i || ids[i] == n_kv, then cell i is not moved - // - std::vector ids(n_kv, n_kv); - - for (uint32_t i0 = 0; i0 < n_used; ++i0) { - const auto & cell0 = kv_self.cells[i0]; - - if (!cell0.is_empty()) { - ids[i0] = i0; - - continue; - } - - // found a hole - fill it with data from the end of the cache - - uint32_t nh = 1; - - // determine the size of the hole - while (i0 + nh < n_used && kv_self.cells[i0 + nh].is_empty()) { - nh++; - } - - uint32_t nf = 0; - uint32_t is = n_kv - 1; - - // starting from the end, find nh non-empty cells - for (; is > i0; --is) { - const auto & cell1 = kv_self.cells[is]; - - if (cell1.is_empty() || ids[is] != n_kv) { - continue; - } - - // non-empty cell which is not yet moved - nf++; - - if (nf == nh) { - break; - } - } - - // this can only happen if `n_used` is not accurate, which would be a bug - GGML_ASSERT(nf == nh && "KV defrag bug: nf != nh"); - - nf = 0; - - uint32_t i1 = is; - - // are we moving a continuous block of memory? - bool cont = false; - - // should we stop searching for the next move? - bool stop = false; - - // go back and move the nf cells to the hole - for (; i1 < n_kv; ++i1) { - auto & cell1 = kv_self.cells[i1]; - - if (cell1.is_empty() || ids[i1] != n_kv) { - if (n_moves == max_moves) { - stop = true; - break; - } - - cont = false; - continue; - } - - // this cell goes to (i0 + nf) - ids[i1] = i0 + nf; - - // move the cell meta data - kv_self.cells[i0 + nf] = cell1; - - // clear the old cell and move the head there - cell1 = llama_kv_cell(); - kv_self.head = n_used; - - if (!cont) { - n_moves++; - cont = true; - } - - nf++; - - if (nf == nh) { - break; - } - } - - if (stop || n_moves == max_moves) { - break; - } - - //LLAMA_LOG_INFO("(tmp log) KV defrag: move [%u, %u) to [%u, %u)\n", is, i1 + 1, i0, i0 + nh); - - i0 += nh - 1; - } - - if (n_moves == 0) { - return; - } - - //LLAMA_LOG_INFO("(tmp log) KV defrag cell moves: %u\n", n_moves); - - //LLAMA_LOG_INFO("expected gf nodes: %u\n", 6*n_moves*n_layer); - -#if 0 - // CPU defrag - // - // TODO: optimizations are possible: - // - multiple threads - // - avoid copying to the host memory when already there - // - // likely not worth the effort, as we have ggml_graph based defrag - // - - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(); - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(); - - const uint32_t kv_size = kv_self.size; - - std::vector buf_k; - std::vector buf_v; - - for (uint32_t il = 0; il < n_layer; ++il) { - const size_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa); - const size_t k_size = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_size); - - const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type); - const size_t v_size = ggml_row_size (kv_self.v_l[il]->type, n_embd_v_gqa*kv_size); - - buf_k.resize(k_size); - buf_v.resize(v_size); - - ggml_backend_tensor_get(kv_self.k_l[il], buf_k.data(), 0, buf_k.size()); - ggml_backend_tensor_get(kv_self.v_l[il], buf_v.data(), 0, buf_v.size()); - - // batch move [i, i+nm) to [id, id+nm) - // note: cells can move only to a lower index - for (uint32_t i = 0; i < n_kv; ++i) { - const uint32_t id = ids[i]; - - if (i == id || id == n_kv) { - continue; - } - - uint32_t nm = 1; - - while (i + nm < n_kv && ids[i + nm] == id + nm) { - nm++; - } - - // move keys - { - const int64_t os = i*k_size_row; - const int64_t od = id*k_size_row; - - memcpy(buf_k.data() + od, buf_k.data() + os, nm*k_size_row); - } - - // move values (note: they are transposed) - { - const int64_t os = i; - const int64_t od = id; - - for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { - memcpy(buf_v.data() + (od + j*kv_size)*v_size_el, buf_v.data() + (os + j*kv_size)*v_size_el, nm*v_size_el); - } - } - - i += nm - 1; - } - - ggml_backend_tensor_set(kv_self.k_l[il], buf_k.data(), 0, buf_k.size()); - ggml_backend_tensor_set(kv_self.v_l[il], buf_v.data(), 0, buf_v.size()); - } -#else - // ggml_graph defrag - - ggml_backend_sched_reset(lctx.sched.get()); - - ggml_cgraph * gf = llama_build_graph_defrag(lctx, ids); - - llama_graph_compute(lctx, gf, lctx.cparams.n_threads, lctx.threadpool); -#endif - - //const int64_t t_end = ggml_time_us(); - - //LLAMA_LOG_INFO("(tmp log) KV defrag time: %.3f ms\n", (t_end - t_start)/1000.0); -} - -static void llama_kv_cache_update_impl(struct llama_context & lctx) { - bool need_reserve = false; - - if (lctx.kv_self.has_shift) { - if (!llama_kv_cache_can_shift(&lctx)) { - GGML_ABORT("The current context does not support K-shift"); - } - - // apply K-shift if needed - if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE) { - ggml_backend_sched_reset(lctx.sched.get()); - - ggml_cgraph * gf = llama_build_graph_k_shift(lctx); - - ggml_backend_sched_alloc_graph(lctx.sched.get(), gf); - - llama_set_k_shift(lctx); - - llama_graph_compute(lctx, gf, lctx.cparams.n_threads, lctx.threadpool); - - need_reserve = true; - } - - { - auto & kv_self = lctx.kv_self; - - kv_self.has_shift = false; - - for (uint32_t i = 0; i < kv_self.size; ++i) { - kv_self.cells[i].delta = 0; - } - } - } - - // defragment the KV cache if needed - if (lctx.kv_self.do_defrag) { - llama_kv_cache_defrag_impl(lctx); - - need_reserve = true; - - lctx.kv_self.do_defrag = false; - } - - // reserve a worst case graph again - if (need_reserve) { - // TODO: extract to a function - // build worst-case graph - uint32_t n_seqs = 1; // TODO: worst-case number of sequences - uint32_t n_tokens = std::min(lctx.cparams.n_ctx, lctx.cparams.n_ubatch); - llama_token token = lctx.model.vocab.token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph - llama_ubatch ubatch = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr}; - ggml_cgraph * gf = llama_build_graph(lctx, ubatch, true); - - // initialize scheduler with the worst-case graph - ggml_backend_sched_reset(lctx.sched.get()); - if (!ggml_backend_sched_reserve(lctx.sched.get(), gf)) { - LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__); - } - } -} - -int32_t llama_set_adapter_lora( - struct llama_context * ctx, - struct llama_adapter_lora * adapter, - float scale) { - ctx->lora[adapter] = scale; - return 0; -} - -int32_t llama_rm_adapter_lora( - struct llama_context * ctx, - struct llama_adapter_lora * adapter) { - auto pos = ctx->lora.find(adapter); - if (pos != ctx->lora.end()) { - ctx->lora.erase(pos); - return 0; - } - - return -1; -} - -void llama_clear_adapter_lora(struct llama_context * ctx) { - ctx->lora.clear(); -} - -int32_t llama_apply_adapter_cvec( - struct llama_context * ctx, - const float * data, - size_t len, - int32_t n_embd, - int32_t il_start, - int32_t il_end) { - return ctx->cvec.apply(ctx->model, data, len, n_embd, il_start, il_end); -} - // // interface implementation // -struct llama_context_params llama_context_default_params() { - struct llama_context_params result = { - /*.n_ctx =*/ 512, - /*.n_batch =*/ 2048, - /*.n_ubatch =*/ 512, - /*.n_seq_max =*/ 1, - /*.n_threads =*/ GGML_DEFAULT_N_THREADS, // TODO: better default - /*.n_threads_batch =*/ GGML_DEFAULT_N_THREADS, - /*.rope_scaling_type =*/ LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED, - /*.pooling_type =*/ LLAMA_POOLING_TYPE_UNSPECIFIED, - /*.attention_type =*/ LLAMA_ATTENTION_TYPE_UNSPECIFIED, - /*.rope_freq_base =*/ 0.0f, - /*.rope_freq_scale =*/ 0.0f, - /*.yarn_ext_factor =*/ -1.0f, - /*.yarn_attn_factor =*/ 1.0f, - /*.yarn_beta_fast =*/ 32.0f, - /*.yarn_beta_slow =*/ 1.0f, - /*.yarn_orig_ctx =*/ 0, - /*.defrag_thold =*/ -1.0f, - /*.cb_eval =*/ nullptr, - /*.cb_eval_user_data =*/ nullptr, - /*.type_k =*/ GGML_TYPE_F16, - /*.type_v =*/ GGML_TYPE_F16, - /*.logits_all =*/ false, - /*.embeddings =*/ false, - /*.offload_kqv =*/ true, - /*.flash_attn =*/ false, - /*.no_perf =*/ true, - /*.abort_callback =*/ nullptr, - /*.abort_callback_data =*/ nullptr, - }; - - return result; -} - struct llama_sampler_chain_params llama_sampler_chain_default_params() { struct llama_sampler_chain_params result = { /*.no_perf =*/ true, @@ -9424,6 +82,57 @@ int64_t llama_time_us(void) { return ggml_time_us(); } +// Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback +static int llama_model_load(const std::string & fname, std::vector & splits, llama_model & model, llama_model_params & params) { + // loading time will be recalculated after the first eval, so + // we take page faults deferred by mmap() into consideration + model.t_load_us = 0; + time_meas tm(model.t_load_us); + + model.t_start_us = tm.t_start_us; + + try { + llama_model_loader ml(fname, splits, params.use_mmap, params.check_tensors, params.kv_overrides); + + ml.print_info(); + + model.hparams.vocab_only = params.vocab_only; + + try { + model.load_arch(ml); + } catch(const std::exception & e) { + throw std::runtime_error("error loading model architecture: " + std::string(e.what())); + } + try { + model.load_hparams(ml); + } catch(const std::exception & e) { + throw std::runtime_error("error loading model hyperparameters: " + std::string(e.what())); + } + try { + model.load_vocab(ml); + } catch(const std::exception & e) { + throw std::runtime_error("error loading model vocabulary: " + std::string(e.what())); + } + + model.load_stats(ml); + model.print_info(); + + if (params.vocab_only) { + LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__); + return 0; + } + + if (!model.load_tensors(ml)) { + return -2; + } + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: error loading model: %s\n", __func__, err.what()); + return -1; + } + + return 0; +} + static struct llama_model * llama_model_load_from_file_impl( const std::string & path_model, std::vector & splits, @@ -9544,460 +253,6 @@ struct llama_model * llama_model_load_from_splits( return llama_model_load_from_file_impl(splits.front(), splits, params); } -struct llama_context * llama_init_from_model( - struct llama_model * model, - struct llama_context_params params) { - - if (!model) { - LLAMA_LOG_ERROR("%s: model cannot be NULL\n", __func__); - return nullptr; - } - - if (params.n_batch == 0 && params.n_ubatch == 0) { - LLAMA_LOG_ERROR("%s: n_batch and n_ubatch cannot both be zero\n", __func__); - return nullptr; - } - - if (params.n_ctx == 0 && model->hparams.n_ctx_train == 0) { - LLAMA_LOG_ERROR("%s: n_ctx and model->hparams.n_ctx_train cannot both be zero\n", __func__); - return nullptr; - } - - if (params.flash_attn && model->arch == LLM_ARCH_GROK) { - LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__); - params.flash_attn = false; - } - - if (params.flash_attn && model->hparams.n_embd_head_k != model->hparams.n_embd_head_v) { - LLAMA_LOG_WARN("%s: flash_attn requires n_embd_head_k == n_embd_head_v - forcing off\n", __func__); - params.flash_attn = false; - } - - if (ggml_is_quantized(params.type_v) && !params.flash_attn) { - LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__); - return nullptr; - } - - llama_context * ctx = new llama_context(*model); - - const auto & hparams = model->hparams; - auto & cparams = ctx->cparams; - - cparams.n_seq_max = std::max(1u, params.n_seq_max); - cparams.n_threads = params.n_threads; - cparams.n_threads_batch = params.n_threads_batch; - cparams.yarn_ext_factor = params.yarn_ext_factor; - cparams.yarn_attn_factor = params.yarn_attn_factor; - cparams.yarn_beta_fast = params.yarn_beta_fast; - cparams.yarn_beta_slow = params.yarn_beta_slow; - cparams.defrag_thold = params.defrag_thold; - cparams.embeddings = params.embeddings; - cparams.offload_kqv = params.offload_kqv; - cparams.flash_attn = params.flash_attn; - cparams.no_perf = params.no_perf; - cparams.pooling_type = params.pooling_type; - - cparams.n_ctx = params.n_ctx == 0 ? hparams.n_ctx_train : params.n_ctx; - cparams.rope_freq_base = params.rope_freq_base == 0.0f ? hparams.rope_freq_base_train : params.rope_freq_base; - cparams.rope_freq_scale = params.rope_freq_scale == 0.0f ? hparams.rope_freq_scale_train : params.rope_freq_scale; - - // this is necessary due to kv_self.n being padded later during inference - cparams.n_ctx = GGML_PAD(cparams.n_ctx, llama_kv_cache_get_padding(cparams)); - - // with causal attention, the batch size is limited by the context size - cparams.n_batch = hparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch; - - // the batch has to be at least GGML_KQ_MASK_PAD because we will be padding the KQ_mask - // this is required by GPU kernels in order to avoid out-of-bounds accesses (e.g. ggml_flash_attn_ext) - // ref: https://github.com/ggerganov/llama.cpp/pull/5021 - if (cparams.n_batch < GGML_KQ_MASK_PAD) { - LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD); - cparams.n_batch = GGML_KQ_MASK_PAD; - } - - cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch); - - cparams.n_ctx_orig_yarn = params.yarn_orig_ctx != 0 ? params.yarn_orig_ctx : - hparams.n_ctx_orig_yarn != 0 ? hparams.n_ctx_orig_yarn : - hparams.n_ctx_train; - - cparams.cb_eval = params.cb_eval; - cparams.cb_eval_user_data = params.cb_eval_user_data; - - auto rope_scaling_type = params.rope_scaling_type; - if (rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED) { - rope_scaling_type = hparams.rope_scaling_type_train; - } - - if (rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_NONE) { - cparams.rope_freq_scale = 1.0f; // never scale if scaling type is none - } - - if (cparams.yarn_ext_factor < 0.0f) { // negative indicates 'not set' - cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_YARN ? 1.0f : 0.0f; - } - - cparams.yarn_attn_factor *= hparams.rope_attn_factor; - - if (cparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) { - if (hparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) { - cparams.pooling_type = LLAMA_POOLING_TYPE_NONE; - } else { - cparams.pooling_type = hparams.pooling_type; - } - } - - if (params.attention_type == LLAMA_ATTENTION_TYPE_UNSPECIFIED) { - cparams.causal_attn = hparams.causal_attn; - } else { - cparams.causal_attn = params.attention_type == LLAMA_ATTENTION_TYPE_CAUSAL; - } - - const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max; - - LLAMA_LOG_INFO("%s: n_seq_max = %u\n", __func__, cparams.n_seq_max); - LLAMA_LOG_INFO("%s: n_ctx = %u\n", __func__, cparams.n_ctx); - LLAMA_LOG_INFO("%s: n_ctx_per_seq = %u\n", __func__, n_ctx_per_seq); - LLAMA_LOG_INFO("%s: n_batch = %u\n", __func__, cparams.n_batch); - LLAMA_LOG_INFO("%s: n_ubatch = %u\n", __func__, cparams.n_ubatch); - LLAMA_LOG_INFO("%s: flash_attn = %d\n", __func__, cparams.flash_attn); - LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, cparams.rope_freq_base); - LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, cparams.rope_freq_scale); - - if (n_ctx_per_seq < hparams.n_ctx_train) { - LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n", - __func__, n_ctx_per_seq, hparams.n_ctx_train); - } - - if (n_ctx_per_seq > hparams.n_ctx_train) { - LLAMA_LOG_WARN("%s: n_ctx_pre_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n", - __func__, n_ctx_per_seq, hparams.n_ctx_train); - } - - ctx->logits_all = params.logits_all; - - // build worst-case graph for encoder if a model contains encoder - ctx->is_encoding = llama_model_has_encoder(model); - - uint32_t kv_size = cparams.n_ctx; - ggml_type type_k = params.type_k; - ggml_type type_v = params.type_v; - - // Mamba only needs a constant number of KV cache cells per sequence - if (llama_model_is_recurrent(model)) { - // Mamba needs at least as many KV cells as there are sequences kept at any time - kv_size = std::max((uint32_t) 1, params.n_seq_max); - // it's probably best to keep as much precision as possible for the states - type_k = GGML_TYPE_F32; // required by ggml_ssm_conv for Mamba's conv_states - type_v = GGML_TYPE_F32; // required by ggml_ssm_scan for Mamba's ssm_states - } - - GGML_ASSERT(hparams.n_embd_head_k % ggml_blck_size(type_k) == 0); - GGML_ASSERT(hparams.n_embd_head_v % ggml_blck_size(type_v) == 0); - - if (!hparams.vocab_only) { - // GPU backends - for (auto * dev : model->devices) { - ggml_backend_t backend = ggml_backend_dev_init(dev, nullptr); - if (backend == nullptr) { - LLAMA_LOG_ERROR("%s: failed to initialize %s backend\n", __func__, ggml_backend_dev_name(dev)); - llama_free(ctx); - return nullptr; - } - ctx->backends.emplace_back(backend); - } - - // add ACCEL backends (such as BLAS) - for (size_t i = 0; i < ggml_backend_dev_count(); ++i) { - ggml_backend_dev_t dev = ggml_backend_dev_get(i); - if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_ACCEL) { - ggml_backend_t backend = ggml_backend_dev_init(dev, nullptr); - if (backend == nullptr) { - LLAMA_LOG_ERROR("%s: failed to initialize %s backend\n", __func__, ggml_backend_dev_name(dev)); - llama_free(ctx); - return nullptr; - } - ctx->backends.emplace_back(backend); - } - } - - // add CPU backend - ctx->backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr); - if (ctx->backend_cpu == nullptr) { - LLAMA_LOG_ERROR("%s: failed to initialize CPU backend\n", __func__); - llama_free(ctx); - return nullptr; - } - ctx->backends.emplace_back(ctx->backend_cpu); - - // create a list of the set_n_threads functions in the backends - for (auto & backend : ctx->backends) { - ggml_backend_dev_t dev = ggml_backend_get_device(backend.get()); - ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr; - if (reg) { - auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads"); - if (ggml_backend_set_n_threads_fn) { - ctx->set_n_threads_fns.emplace_back(backend.get(), ggml_backend_set_n_threads_fn); - } - } - } - - llama_set_abort_callback(ctx, params.abort_callback, params.abort_callback_data); - - if (!llama_kv_cache_init(ctx->kv_self, ctx->model, ctx->cparams, type_k, type_v, kv_size, cparams.offload_kqv)) { - LLAMA_LOG_ERROR("%s: llama_kv_cache_init() failed for self-attention cache\n", __func__); - llama_free(ctx); - return nullptr; - } - - { - size_t memory_size_k = 0; - size_t memory_size_v = 0; - - for (auto & k : ctx->kv_self.k_l) { - memory_size_k += ggml_nbytes(k); - } - - for (auto & v : ctx->kv_self.v_l) { - memory_size_v += ggml_nbytes(v); - } - - LLAMA_LOG_INFO("%s: KV self size = %7.2f MiB, K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__, - (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f), - ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f), - ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f)); - } - - // graph outputs buffer - { - // resized during inference when a batch uses more outputs - if (llama_output_reserve(*ctx, params.n_seq_max) < params.n_seq_max) { - LLAMA_LOG_ERROR("%s: failed to reserve initial output buffer\n", __func__); - llama_free(ctx); - return nullptr; - } - - LLAMA_LOG_INFO("%s: %10s output buffer size = %8.2f MiB\n", __func__, - ggml_backend_buffer_name(ctx->buf_output.get()), - ggml_backend_buffer_get_size(ctx->buf_output.get()) / 1024.0 / 1024.0); - } - - // scheduler and compute buffers - { - // buffer types used for the compute buffer of each backend - std::vector backend_buft; - std::vector backend_ptrs; - for (auto & backend : ctx->backends) { - auto * buft = ggml_backend_get_default_buffer_type(backend.get()); - auto backend_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get())); - if (backend_type == GGML_BACKEND_DEVICE_TYPE_CPU && !model->devices.empty()) { - // use the host buffer of the first device CPU for faster transfer of the intermediate state - auto * dev = model->devices[0]; - auto * host_buft = ggml_backend_dev_host_buffer_type(dev); - if (host_buft) { - buft = host_buft; - } - } - backend_buft.push_back(buft); - backend_ptrs.push_back(backend.get()); - } - - const size_t max_nodes = model->max_nodes(); - - // buffer used to store the computation graph and the tensor meta data - ctx->buf_compute_meta.resize(ggml_tensor_overhead()*max_nodes + ggml_graph_overhead_custom(max_nodes, false)); - - // TODO: move these checks to ggml_backend_sched - // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary - bool pipeline_parallel = - model->n_devices() > 1 && - model->params.n_gpu_layers > (int)model->hparams.n_layer && - model->params.split_mode == LLAMA_SPLIT_MODE_LAYER && - params.offload_kqv; - - // pipeline parallelism requires support for async compute and events in all devices - if (pipeline_parallel) { - for (auto & backend : ctx->backends) { - auto dev_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get())); - if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU) { - // ignore CPU backend - continue; - } - auto * dev = ggml_backend_get_device(backend.get()); - ggml_backend_dev_props props; - ggml_backend_dev_get_props(dev, &props); - if (!props.caps.async || !props.caps.events) { - // device does not support async compute or events - pipeline_parallel = false; - break; - } - } - } - - ctx->sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, pipeline_parallel)); - - if (pipeline_parallel) { - LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(ctx->sched.get())); - } - - // initialize scheduler with the worst-case graph - uint32_t n_seqs = 1; // TODO: worst-case number of sequences - uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch); - llama_token token = ctx->model.vocab.token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph - - llama_ubatch ubatch_pp = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr}; - ggml_cgraph * gf_pp = llama_build_graph(*ctx, ubatch_pp, true); - - // reserve pp graph first so that buffers are only allocated once - ggml_backend_sched_reserve(ctx->sched.get(), gf_pp); - int n_splits_pp = ggml_backend_sched_get_n_splits(ctx->sched.get()); - int n_nodes_pp = ggml_graph_n_nodes(gf_pp); - - // reserve with tg graph to get the number of splits and nodes - llama_ubatch ubatch_tg = { true, 1, 1, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr}; - ggml_cgraph * gf_tg = llama_build_graph(*ctx, ubatch_tg, true); - ggml_backend_sched_reserve(ctx->sched.get(), gf_tg); - int n_splits_tg = ggml_backend_sched_get_n_splits(ctx->sched.get()); - int n_nodes_tg = ggml_graph_n_nodes(gf_tg); - - // reserve again with pp graph to avoid ggml-alloc reallocations during inference - gf_pp = llama_build_graph(*ctx, ubatch_pp, true); - if (!ggml_backend_sched_reserve(ctx->sched.get(), gf_pp)) { - LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__); - llama_free(ctx); - return nullptr; - } - - for (size_t i = 0; i < backend_ptrs.size(); ++i) { - ggml_backend_t backend = backend_ptrs[i]; - ggml_backend_buffer_type_t buft = backend_buft[i]; - size_t size = ggml_backend_sched_get_buffer_size(ctx->sched.get(), backend); - if (size > 1) { - LLAMA_LOG_INFO("%s: %10s compute buffer size = %8.2f MiB\n", __func__, - ggml_backend_buft_name(buft), - size / 1024.0 / 1024.0); - } - } - - if (n_nodes_pp == n_nodes_tg) { - LLAMA_LOG_INFO("%s: graph nodes = %d\n", __func__, n_nodes_pp); - } else { - LLAMA_LOG_INFO("%s: graph nodes = %d (with bs=%d), %d (with bs=1)\n", __func__, n_nodes_pp, n_tokens, n_nodes_tg); - } - if (n_splits_pp == n_splits_tg) { - LLAMA_LOG_INFO("%s: graph splits = %d\n", __func__, n_splits_pp); - } else { - LLAMA_LOG_INFO("%s: graph splits = %d (with bs=%d), %d (with bs=1)\n", __func__, n_splits_pp, n_tokens, n_splits_tg); - } - } - } - - return ctx; -} - -struct llama_context * llama_new_context_with_model( - struct llama_model * model, - struct llama_context_params params) { - return llama_init_from_model(model, params); -} - -// -// kv cache -// - -// TODO: tmp bridges below until `struct llama_kv_cache` is exposed through the public API - -struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_seq_max) { - return llama_kv_cache_view_init(ctx->kv_self, n_seq_max); -} - -void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view) { - llama_kv_cache_view_update(view, ctx->kv_self); -} - -int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx) { - return llama_get_kv_cache_token_count(ctx->kv_self); -} - -int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx) { - return llama_get_kv_cache_used_cells(ctx->kv_self); -} - -void llama_kv_cache_clear(struct llama_context * ctx) { - llama_kv_cache_clear(ctx->kv_self); -} - -bool llama_kv_cache_seq_rm(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1) { - return llama_kv_cache_seq_rm(ctx->kv_self, seq_id, p0, p1); -} - -void llama_kv_cache_seq_cp(struct llama_context * ctx, llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) { - if (seq_id_src == seq_id_dst) { - return; - } - llama_kv_cache_seq_cp(ctx->kv_self, seq_id_src, seq_id_dst, p0, p1); -} - -void llama_kv_cache_seq_keep(struct llama_context * ctx, llama_seq_id seq_id) { - llama_kv_cache_seq_keep(ctx->kv_self, seq_id); -} - -void llama_kv_cache_seq_add(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) { - if (delta == 0) { - return; - } - - llama_kv_cache_seq_add(ctx->kv_self, seq_id, p0, p1, delta); -} - -void llama_kv_cache_seq_div(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) { - if (d == 1) { - return; - } - - llama_kv_cache_seq_div(ctx->kv_self, seq_id, p0, p1, d); -} - -llama_pos llama_kv_cache_seq_pos_max(struct llama_context * ctx, llama_seq_id seq_id) { - return llama_kv_cache_seq_pos_max(ctx->kv_self, seq_id); -} - -void llama_kv_cache_defrag(struct llama_context * ctx) { - llama_kv_cache_defrag(ctx->kv_self); -} - -void llama_kv_cache_update(struct llama_context * ctx) { - llama_kv_cache_update_impl(*ctx); -} - -bool llama_kv_cache_can_shift(struct llama_context * ctx) { - return llama_kv_cache_can_shift(ctx->kv_self); -} - -/// - -int32_t llama_encode( - struct llama_context * ctx, - struct llama_batch batch) { - const int ret = llama_encode_impl(*ctx, batch); - if (ret != 0) { - LLAMA_LOG_ERROR("%s: failed to encode, ret = %d\n", __func__, ret); - } - - return ret; -} - -int32_t llama_decode( - struct llama_context * ctx, - struct llama_batch batch) { - const int ret = llama_decode_impl(*ctx, batch); - if (ret != 0) { - LLAMA_LOG_ERROR("%s: failed to decode, ret = %d\n", __func__, ret); - } - - return ret; -} - // // chat templates // @@ -10065,7 +320,6 @@ const char * llama_print_system_info(void) { static std::string s; s.clear(); // Clear the string, since it's static, otherwise it will accumulate data from previous calls. - for (size_t i = 0; i < ggml_backend_reg_count(); i++) { auto * reg = ggml_backend_reg_get(i); auto * get_features_fn = (ggml_backend_get_features_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_get_features"); @@ -10084,43 +338,3 @@ const char * llama_print_system_info(void) { return s.c_str(); } - -// -// perf -// - -struct llama_perf_context_data llama_perf_context(const struct llama_context * ctx) { - struct llama_perf_context_data data = {}; - - if (ctx == nullptr) { - return data; - } - - data.t_start_ms = 1e-3 * ctx->t_start_us; - data.t_load_ms = 1e-3 * ctx->t_load_us; - data.t_p_eval_ms = 1e-3 * ctx->t_p_eval_us; - data.t_eval_ms = 1e-3 * ctx->t_eval_us; - data.n_p_eval = std::max(1, ctx->n_p_eval); - data.n_eval = std::max(1, ctx->n_eval); - - return data; -} - -void llama_perf_context_print(const struct llama_context * ctx) { - const auto data = llama_perf_context(ctx); - - const double t_end_ms = 1e-3 * ggml_time_us(); - - LLAMA_LOG_INFO("%s: load time = %10.2f ms\n", __func__, data.t_load_ms); - LLAMA_LOG_INFO("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n", - __func__, data.t_p_eval_ms, data.n_p_eval, data.t_p_eval_ms / data.n_p_eval, 1e3 / data.t_p_eval_ms * data.n_p_eval); - LLAMA_LOG_INFO("%s: eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n", - __func__, data.t_eval_ms, data.n_eval, data.t_eval_ms / data.n_eval, 1e3 / data.t_eval_ms * data.n_eval); - LLAMA_LOG_INFO("%s: total time = %10.2f ms / %5d tokens\n", __func__, (t_end_ms - data.t_start_ms), (data.n_p_eval + data.n_eval)); -} - -void llama_perf_context_reset(struct llama_context * ctx) { - ctx->t_start_us = ggml_time_us(); - ctx->t_eval_us = ctx->n_eval = 0; - ctx->t_p_eval_us = ctx->n_p_eval = 0; -} diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index b4e3631ed0..c86ffb64e9 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -4113,7 +4113,7 @@ static std::vector> make_test_cases_eval() { for (int n_mats : {4, 8}) { for (int n_used : {1, 2, 4}) { for (bool b : {false, true}) { - for (int n : {1, 32}) { + for (int n : {1, 32, 129}) { int m = 512; int k = 256; test_cases.emplace_back(new test_mul_mat_id(type_a, type_b, n_mats, n_used, b, m, n, k)); diff --git a/tests/test-chat.cpp b/tests/test-chat.cpp index 35a307c632..a1034b1a41 100644 --- a/tests/test-chat.cpp +++ b/tests/test-chat.cpp @@ -480,6 +480,21 @@ static void test_msgs_oaicompat_json_conversion() { "]" ), common_chat_msgs_to_json_oaicompat({message_assist_call_python}).dump(2)); + + auto res = common_chat_msgs_parse_oaicompat(json::parse("[{\"role\": \"assistant\", \"tool_calls\": []}]")); + assert_equals(1, res.size()); + assert_equals(res[0].role, "assistant"); + assert_equals(true, res[0].content.empty()); + assert_equals(true, res[0].tool_calls.empty()); + + try { + common_chat_msgs_parse_oaicompat(json::parse("[{\"role\": \"assistant\"}]")); + throw std::runtime_error("Expected exception"); + } catch (const std::exception & e) { + if (std::string(e.what()).find("'content'") == std::string::npos) { + throw std::runtime_error("Expected exception about missing 'content'"); + } + } } static void test_tools_oaicompat_json_conversion() { @@ -751,6 +766,19 @@ static void test_template_output_parsers() { "{\n \"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}", COMMON_CHAT_FORMAT_HERMES_2_PRO)); + assert_msg_equals(message_assist_thoughts_unparsed_think, + common_chat_parse("I'm thinkingHello, world!\nWhat's up?", + COMMON_CHAT_FORMAT_HERMES_2_PRO)); + assert_msg_equals(message_assist_thoughts_unparsed_think, + common_chat_parse("I'm thinkingHello, world!\nWhat's up?", + COMMON_CHAT_FORMAT_HERMES_2_PRO)); + assert_msg_equals(message_assist_thoughts, + common_chat_parse("I'm thinkingHello, world!\nWhat's up?", + COMMON_CHAT_FORMAT_HERMES_2_PRO_EXTRACT_REASONING)); + assert_msg_equals(message_assist_thoughts, + common_chat_parse("I'm thinkingHello, world!\nWhat's up?", + COMMON_CHAT_FORMAT_HERMES_2_PRO_EXTRACT_REASONING)); + test_templates(tmpls.get(), end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false); test_templates(tmpls.get(), end_tokens, message_assist_call, tools, "\n" diff --git a/tests/test-quantize-fns.cpp b/tests/test-quantize-fns.cpp index c77c8ed138..037c0582bb 100644 --- a/tests/test-quantize-fns.cpp +++ b/tests/test-quantize-fns.cpp @@ -120,13 +120,7 @@ int main(int argc, char * argv[]) { generate_data(0.0, test_data.size(), test_data.data()); generate_data(1.0, test_data2.size(), test_data2.data()); - // Initialize GGML, ensures float conversion tables are initialized - struct ggml_init_params ggml_params = { - /* .mem_size = */ 1*1024, - /* .mem_buffer = */ NULL, - /* .no_alloc = */ true, - }; - struct ggml_context * ctx = ggml_init(ggml_params); + ggml_cpu_init(); int num_failed = 0; bool failed = false; @@ -188,7 +182,5 @@ int main(int argc, char * argv[]) { printf("%d tests failed\n", num_failed); } - ggml_free(ctx); - return num_failed > 0; }