happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
7,882 Commits 542 Branches 5,540 Tags 1.4 GiB
Commit Graph

51 Commits

Author SHA1 Message Date
Yu, Zijun dc77cbb3f6 STYLE: make get_types_to_requant a function 2026-01-15 11:20:31 -08:00
Yu, Zijun 2ad1147b9b Improve debug util; Eliminate nop ReshapeReshape 2026-01-15 11:20:31 -08:00
Yu, Zijun 6926655f5b Add custom quant type: q8_1_c, q4_0_128 2026-01-15 11:20:31 -08:00
Yu, Zijun b593428eb3 Dequantize q4_1 q4_k q6_k for NPU 2026-01-15 11:20:31 -08:00
Yu, Zijun 9900245e0b Fix test-backend-ops: Treat quantized tensors as weights 2026-01-15 11:20:31 -08:00
Yu, Zijun 65e1b1af6d Fix after rebasing 
- Layout of cache k and cache v are unified: [seq, n_head, head_size]
- Add CPY and FLASH_ATTN_EXT, flash attn is not used yet
- Skip test-backend-ops due to flash attn test crash
- Add mutex around graph conversion to avoid test-thread-safety fali in the future
- Update NPU config
- Update GPU config to disable SDPA opt to make phi-3 run
 2026-01-15 11:19:15 -08:00
Yu, Zijun f4123be967 Fix test-backend-ops 2026-01-15 11:19:15 -08:00
Yu, Zijun 7bda5021f9 Fix NPU 2026-01-15 11:19:15 -08:00
Yu, Zijun 37ff226bb6 Use CiD for NPU 2026-01-15 11:19:15 -08:00
Yu, Zijun fc865340d5 Fix test-backend-ops 2026-01-15 10:26:28 -08:00
Yu, Zijun 4e7f04a307 Fix llama-perplexity 2026-01-15 10:26:28 -08:00
Yu, Zijun 6dc4b90635 Fix NPU 2026-01-15 10:26:28 -08:00
Yu, Zijun 44f4cf34b1 Fix Phi3 ROPE; Add test-backend-ops 2026-01-15 10:26:28 -08:00
Yu, Zijun f3c0519096 Reduce memory: free ov weights node after graph conversion 2026-01-15 10:20:18 -08:00
Yu, Zijun ebc4fc9f95 Fuse to SDPA 2026-01-15 10:20:18 -08:00
Yu, Zijun 4c582ac7a3 Statful transformation for CPU GPU 2026-01-15 10:20:18 -08:00
Yu, Zijun 8afee795ad Update clang-format 2026-01-15 10:20:18 -08:00
Yu, Zijun 593484ce5f Refactor: clean, fix warning 2026-01-15 10:20:18 -08:00
Yu, Zijun 592d7f8bbb Change due to ggml cgraph changes, llama-3.2 CPU work 2026-01-15 10:20:18 -08:00
Yu, Zijun d9ca8f5dbe NPU support version 2: prefill + kvcache 2026-01-15 10:20:18 -08:00
Yu, Zijun 34531abce4 draft NPU support version 2: prefill + kvcache 2026-01-15 10:20:18 -08:00
Yu, Zijun 7fec223334 Add initial NPU support 2026-01-15 10:20:18 -08:00
Yu, Zijun 8ac5c225aa FIX: set_max_token_len 2026-01-15 10:20:18 -08:00
Yu, Zijun 0d505b4e56 STYLE and minor REFACTOR 2026-01-15 10:10:00 -08:00
Yu, Zijun 0d009fe61a FEAT: Add all conversion code from ov side 2026-01-15 10:10:00 -08:00
Viraj Wadhwa ffabe95e2a Rebase - Bring up to date and fix build process 2026-01-15 10:09:23 -08:00
Yu, Zijun a8e5efa44e PERF: compile once (dynamic graph + cache) 2026-01-15 10:05:41 -08:00
Yu, Zijun 7d5e234254 FEAT: improve debug capability 2026-01-15 10:05:41 -08:00
Yu, Zijun 0a8cc9ab03 BUILD: update build doc, add cmake preset, add CACHE_DIR env var 2026-01-15 10:05:41 -08:00
Yu, Zijun c04966cda6 REFACTOR: support weigts as constant 2026-01-15 10:05:41 -08:00
Yu, Zijun 91d2a195b5 change op mappings to list in openvino_supports_op 2026-01-15 10:05:41 -08:00
Yu, Zijun 651b2c06cb * Use find_package in CMake to configure OpenVINO 
* Remove OPENVINO_OP_DEBUG
* Simplify set_input_output in decoder
* Fix CPY in set_input_output
* Use params from converted ov model in setting input
 2026-01-15 10:05:41 -08:00
zhanmyz 84be5c6f15 1. Delete some comments 
2. Process Prompt and predict first token is OK
 2026-01-15 10:05:41 -08:00
zhanmyz eac9a99530 1. Solve the AC issue of Permute+VIEW and MULMAL issue in the phase of “1. Process Prompt and predict the first token”. 
2. There is still an AC issue in the "2. Predict the subsequent tokens phase" and it is being debugged.
   A deviation has been detected in the computation of OpenVINO's CPY Node at stage 2, and it is currently being fixed.
 2026-01-15 10:05:41 -08:00
zhanmyz 8ae700ae11 Process Prompt and predict first token is OK 2026-01-15 10:05:41 -08:00
zhanmyz 8020138406 add debug info 2026-01-15 10:05:41 -08:00
zhanmyz b02265a507 1. In the Prompt process and predict first token stage, the PERMUTE node needs to be integrated into the OV Frontend 
2. In the predict latest token stage, the VIEW, CONT, Reshape need to be integrated into the OV Frontend.
 2026-01-15 10:05:41 -08:00
zhanmyz 467a5ddf04 1. Update the implementation of CPY node when it's non-contiguous 
2. Remove duplicate get node operation function
 2026-01-15 10:05:41 -08:00
zhanmyz cff473a9e2 1. All operators implemented using OpenVINO can be successfully executed individually. 
2. VIEW op output tensor shape is not same with CONT(non-contiguous) input tensor shape
3. CPY(non-contiguous) can't be implemented with original input/output tensor shape and data(need change the original shape when create input/output tensor)

Currently. VIEW op executed in the ggml backend and others executed in the OpenVINO Frontend.
 2026-01-15 10:05:41 -08:00
zhanmyz e08a7fda33 All adjacent ops can conversion but calculation result is wrong and need debugging 2026-01-15 10:05:41 -08:00
zhanmyz 9a7b7d8d6d OV Frontend supports GET_ROWS/RMS_NORM/MUL/MUL_MAT/ROPE/SCALE/SOFTMAX/ADD adjacent op graph conversion 2026-01-15 10:05:41 -08:00
zhanmyz 95ae982d59 OV Frontend supports GET_ROWS/RMS_NORM/MUL/MUL_MAT graph conversion of consecutive OPs 2026-01-15 10:05:41 -08:00
zhanmyz 901f7347ff Execute CONT & VIEW operators in OV Frontend is OK 2026-01-15 10:05:41 -08:00
zhanmyz afb8594194 add tmp source code files 2026-01-15 10:05:41 -08:00
zhanmyz 8484769981 add implementation of MUL_MAT, CPY, CONT of GGML ops using OV ops 2026-01-15 10:05:41 -08:00
zhanmyz 2b04bd43be Add MUL_MAT,CPY,CONT as operators implemented in OpenVINO for GGML backend 2026-01-15 10:05:41 -08:00
yumengbo 590f587b27 Add support for UNARY SILU op . Fix pytorch impl bugs. 2026-01-15 10:05:41 -08:00
yumengbo b100f89bad Change to implementation following pytorch frontend 2026-01-15 10:05:41 -08:00
yumengbo e95f29cbc0 Fix issue for output memory copy of infer request 2026-01-15 10:05:41 -08:00
yumengbo 5b46dc23be Change output for infer request to set output tensor. Support scale, view op. 2026-01-15 10:05:41 -08:00