* examples/finetune -opt SGD (stochastic gradient descent) memory opt
add unit tested GGML_OPT_OPTIMIZER_SGD to ggml - avoids allocating
m, v tensors.
support finetune.cpp arg -opt SGD (or sgd). (default adamw as before)
llama 3.2-1b-F32 result: observed 11gb gpu ram (41 sec/epoch)
when using SGD instead of 19gb (55 sec/epoch) using adamw.
(wikipedia 100 lines finetune)
(
using the same GPU memory, adamw can only do before OOM 512
batch/context, reaching:
train: [███████▉] data=0000140/0000140 loss=0.02575±0.00099 acc=99.52±0.03% t=00:00:47 ETA=00:00:00
val: [███████▉] data=0000008/0000008 loss=4.76565±0.28810 acc=41.46±0.77% t=00:00:00 ETA=00:00:00
SGD is superior, though it converges slower, with max before OOM 1728
batch/context (esp see the better validation perf):
train: [███████▉] data=0000039/0000039 loss=0.00371±0.00010 acc=99.96±0.01% t=00:00:41 ETA=00:00:00
val: [███████▉] data=0000003/0000003 loss=5.11406±0.76034 acc=48.01±0.69% t=00:00:01 ETA=00:00:00
)
note: when finetuning long enough (or w/ enough -lr),
validation accuracy *eventually* drops ('catastrophic forgetting')
-lr-half (halflife) option useful for SGD to avoid oscillation or
super slow underdamped learning (makes setting -lr more forgiving).
terminal -lr for now is set by lr-halvings i.e. if you want at most
1/8 the inital -lr you set -lr-halvings 3.
note: objective loss not directly comparable between adamw, sgd? -
check perplexity or accuracy or consider relative improvements
for convergence
new finetune args -wd 1e-9 to enable weight decay in sgd or adamw,
and max -epochs N (default 2 as before)
cache (1 - wd*alpha) in 'adamw' opt struct -
no noticeable perf benefit, disabled (still done
for new SGD though)
since opt. memory is pre-allocated, the ggml_opt_get_optimizer_params
would probably be able to change between SGD and AdamW with each epoch
but would need to use adamw for the first (unconfirmed - no cmdline arg
to set such a policy yet)
test-opt checks adamw as before and now sgd (except for a few disabled
tests for sgd only; probably just needs logging values and adding
alternate reference values); tolerance on the 'regression'
test is broader for sgd (so we don't need many more epochs)
* Vulkan: Implement GGML_OP_OPT_STEP_SGD
* tests: Fix OPT_STEP_SGD test-backend-ops
* SGD op param store weight-decay and not 1-alpha*wd
* minor + cosmetic changes
* fix vulkan sgd
* try CI fix
---------
Co-authored-by: 0cc4m <picard12@live.de>
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Checkpoint from VS Code for coding agent session
* Initial plan
* Fix typo in --override-tensor-draft flag implementation
* Add null termination for speculative tensor buffer overrides
* Apply suggestions from code review
* Apply suggestions from code review
* Extract tensor override parsing logic to common function (addresses @slaren's feedback)
* Apply suggestions from code review
* Apply suggestions
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Diego Devesa <slarengh@gmail.com>
* llama-server : implement universal assisted decoding
* Erase prompt tail for kv-cache
* set vocab_dft_compatible in common_speculative
* rename ctx_main to ctx_tgt
* move vocab_dft_compatible to spec struct
* clear mem_dft, remove mem
* detokenize id_last for incompatible models
* update comment
* add --spec-replace flag
* accept special tokens when translating between draft/main models
* Escape spec-replace
* clamp draft result to size to params.n_draft
* fix comment
* clean up code
* restore old example
* log common_speculative_are_compatible in speculative example
* fix
* Update common/speculative.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Update common/speculative.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Update common/speculative.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Add support for Llada-8b: diffusion model
* Add README
* Fix README and convert_hf_to_gguf
* convert_hf_to_gguf.py: address review comments
* Make everything in a single example
* Remove model-specific sampling
* Remove unused argmax
* Remove braced initializers, improve README.md a bit
* Add diffusion specific gguf params in set_vocab, remove setting rope_theta and rms_norm_eps
* Remove adding the mask token
* Move add_add_bos_token to set_vocab
* use add_bool in gguf_writer.py
* imatrix : allow processing multiple chunks per batch
* perplexity : simplify filling the batch
* imatrix : fix segfault when using a single chunk per batch
* imatrix : use GGUF to store imatrix data
* imatrix : fix conversion problems
* imatrix : use FMA and sort tensor names
* py : add requirements for legacy imatrix convert script
* perplexity : revert changes
* py : include imatrix converter requirements in toplevel requirements
* imatrix : avoid using designated initializers in C++
* imatrix : remove unused n_entries
* imatrix : allow loading mis-ordered tensors
Sums and counts tensors no longer need to be consecutive.
* imatrix : more sanity checks when loading multiple imatrix files
* imatrix : use ggml_format_name instead of std::string concatenation
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
* quantize : use unused imatrix chunk_size with LLAMA_TRACE
* common : use GGUF for imatrix output by default
* imatrix : two-way conversion between old format and GGUF
* convert : remove imatrix to gguf python script
* imatrix : use the function name in more error messages
* imatrix : don't use FMA explicitly
This should make comparisons between the formats easier
because this matches the behavior of the previous version.
* imatrix : avoid returning from void function save_imatrix
* imatrix : support 3d tensors with MUL_MAT
* quantize : fix dataset name loading from gguf imatrix
* common : move string_remove_suffix from quantize and imatrix
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* imatrix : add warning when legacy format is written
* imatrix : warn when writing partial data, to help guess dataset coverage
Also make the legacy format store partial data
by using neutral values for missing data.
This matches what is done at read-time for the new format,
and so should get the same quality in case the old format is still used.
* imatrix : avoid loading model to convert or combine imatrix
* imatrix : avoid using imatrix.dat in README
---------
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Support diffusion models: Add Dream 7B
* Move diffusion to examples
* Move stuff to examples. Add patch to not use kv-cache
* Address review comments
* Make sampling fast
* llama: remove diffusion functions
* Add basic timings + cleanup
* More cleanup
* Review comments: better formating, use LOG instead std::cerr, re-use batch, use ubatch instead of max_length
* fixup!
* Review: move everything to diffusion-cli for now
* initial commit for handling extra template kwargs
* enable_thinking and assistant prefill cannot be enabled at the same time
* can set chat_template_kwargs in command line
* added doc
* fixed formatting
* add support for extra context in generic template init
* coding standard: common/chat.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* coding standard: common/chat.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Apply suggestions from code review
coding standard: cosmetic changes
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* fix merge conflict
* chat.cpp: simplify calls to apply to ensure systematic propagation of extra_context (+ the odd existing additional_context)
* normalize environment variable name
* simplify code
* prefill cannot be used with thinking models
* compatibility with the new reasoning-budget parameter
* fix prefill for non thinking models
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Olivier Chafik <olivier.chafik@gmail.com>
* llama : add thread safety test
* llamafile : remove global state
* llama : better LLAMA_SPLIT_MODE_NONE logic
when main_gpu < 0 GPU devices are not used
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Currently when a model generates output which looks like a tool call,
but is invalid an exception is thrown and not handled, causing the cli
or llama-server to bail. Instead, handle the chat parser exception and
simply return the generated text in such cases.
Signed-off-by: Piotr Stankiewicz <piotr.stankiewicz@docker.com>
* cmake: Simplify build-info.cpp generation
The rebuild of build-info.cpp still gets triggered when .git/index gets
changes.
* cmake: generate build-info.cpp in build dir
* llama : deprecate llama_kv_self_ API
ggml-ci
* llama : allow llama_memory_(nullptr)
ggml-ci
* memory : add flag for optional data clear in llama_memory_clear
ggml-ci
* threading: support for GGML_SCHED_PRIO_LOW, update thread info on Windows to avoid throttling
We talked about adding LOW priority for GGML threads in the original threadpool PR.
It might be useful for some cases to avoid contention.
Latest Windows ARM64 releases started parking (offlining) the CPU cores
more aggresively which results in suboptimal performance with n_threads > 4.
To deal with that we now disable Power Throttling for our threads for the NORMAL
and higher priorities.
Co-authored-by: Diego Devesa <slarengh@gmail.com>
* threading: disable SetThreadInfo() calls for older Windows versions
* Update tools/llama-bench/llama-bench.cpp
Co-authored-by: Diego Devesa <slarengh@gmail.com>
---------
Co-authored-by: Diego Devesa <slarengh@gmail.com>
* convert: add support for BertForSequenceClassification
* add support for reranking using BertForSequenceClassification
* merge checks of eos and sep
* fix lint
---------
Co-authored-by: dinhhuy <huy.dinh@brains-tech.co.jp>
* add preludes to content on partial regex match
* allow all parsers to parse non-tool-call content.
* tweak order of <|python_tag|> vs <function= parsing for functionary v3.1 format. still not ideal but hopefully less prone to crash
* fix deltas of tool_call.function.name
* fix tool_call.id (was in tool_call.function.id!) + add function type
* add tool_call.type
* populate empty tool_call.function.arguments on first delta
* convert ok, load ok
* warmup ok
* test
* still does not work?
* fix padding
* temporary give up
* fix merge conflict
* build_ultravox()
* rm test
* fix merge conflict
* add necessary mtmd APIs
* first working version (only 4s of audio)
* will this monster compile?
* fix compile
* please compile
* fPIC
* fix windows
* various fixes
* clean up audio_helpers
* fix conversion
* add some debug stuff
* long audio input ok
* adapt the api
* add --audio arg
* final touch UX
* add miniaudio to readme
* fix typo
* refactor kv metadata
* mtmd_default_marker()
* llama/ggml: add LLM training support
more compact progress bar
llama_save_model_to_file
llama_opt_param_filter
ggml_graph_dup force_grads
refactor ggml_opt, fix test-opt
* remove logits_all
* refactor CUDA implementation for ACC
* reset graph at beginning of opt period
* server : (experimental) vision support via libmtmd
* mtmd : add more api around mtmd_image_tokens
* mtmd : add more api around mtmd_image_tokens
* mtmd : ability to calc image hash
* shared_ptr for mtmd_image_tokens
* move hash to user-define ID (fixed)
* abstract out the batch management
* small fix
* refactor logic adding tokens to batch
* implement hashing image
* use FNV hash, now hash bitmap instead of file data
* allow decoding image embedding to be split into batches
* rm whitespace
* disable some features when mtmd is on
* fix --no-mmproj-offload
* mtmd_context_params no timings
* refactor server_inp to server_tokens
* fix the failing test case
* init
* wip
* working version
* add mtmd::bitmaps
* add test target
* rm redundant define
* test: mtmd_input_chunks_free
* rm outdated comment
* fix merging issue
* explicitly create mtmd::input_chunks
* mtmd_input_chunk_copy
* add clone()
* improve server_input struct
* clip : fix confused naming ffn_up and ffn_down
* rm ffn_i/o/g naming
* rename n_embd, n_ff
* small fix
* no check n_ff
* fix detokenize
* add const to various places
* add warning about breaking changes
* add c api
* helper: use mtmd_image_tokens_get_n_pos
* fix ctx_shift
* fix name shadowing
* more strict condition
* support remote image_url
* remote image_url log
* add CI test
* do not log base64
* add "has_multimodal" to /props
* remove dangling image
* speculative: use slot.cache_tokens.insert
* Apply suggestions from code review
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* rm can_be_detokenized
* on prmpt processing done, assert cache_tokens.size
* handle_completions_impl returns void
* adapt the new web ui
* update docs and hot topics
* rm assert
* small fix (2)
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* rework the input area
* process selected file
* change all icons to heroicons
* fix thought process collapse
* move conversation more menu to sidebar
* sun icon --> moon icon
* rm default system message
* stricter upload file check, only allow image if server has mtmd
* build it
* add renaming
* better autoscroll
* build
* add conversation group
* fix scroll
* extra context first, then user input in the end
* fix <hr> tag
* clean up a bit
* build
* add mb-3 for <pre>
* throttle adjustTextareaHeight to make it less laggy
* (nits) missing padding in sidebar
* rm stray console log
* cmake : do not include ./src as public for libllama
ggml-ci
* cmake : rework tests
ggml-ci
* llguidance : remove unicode include
ggml-ci
* cmake : make c++17 private
ggml-ci
* arg : clean up handling --mmproj with -hf
* rm change about no_mmproj
* Revert "rm change about no_mmproj"
This reverts commit 2cac8e0efb.
* handle no_mmproj explicitly
* skip download mmproj on examples not using it
* mtmd : merge `llava-cli` and `gemma3-cli` into single `mtmd-cli`
* support for minicpmv
* remove cpp files of llava and minicpmv
* update hot topics
* mtmd : add not supported msg for qwen2vl
* Update examples/llava/mtmd.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* common: custom hf endpoint support
Add support for custom huggingface endpoints via HF_ENDPOINT environment variable
You can now specify a custom huggingface endpoint using the HF_ENDPOINT environment variable when using the --hf-repo flag, which works similarly to huggingface-cli's endpoint configuration.
Example usage:
HF_ENDPOINT=https://hf-mirror.com/ ./bin/llama-cli --hf-repo Qwen/Qwen1.5-0.5B-Chat-GGUF --hf-file qwen1_5-0_5b-chat-q2_k.gguf -p "The meaning to life and the universe is"
The trailing slash in the URL is optional:
HF_ENDPOINT=https://hf-mirror.com ./bin/llama-cli --hf-repo Qwen/Qwen1.5-0.5B-Chat-GGUF --hf-file qwen1_5-0_5b-chat-q2_k.gguf -p "The meaning to life and the universe is"
* Update common/arg.cpp
readability Improvement
Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com>
* Apply suggestions from code review
---------
Co-authored-by: ベアトリーチェ <148695646+MakiSonomura@users.noreply.github.com>
Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com>
* (wip) refactor downloading system [no ci]
* fix all examples
* fix mmproj with -hf
* gemma3: update readme
* only handle mmproj in llava example
* fix multi-shard download
* windows: fix problem with std::min and std::max
* fix 2
* server : Bump cpp-httplib to include AF_UNIX windows support
Signed-off-by: Piotr Stankiewicz <piotr.stankiewicz@docker.com>
* server : Allow running the server example on a unix socket
Signed-off-by: Piotr Stankiewicz <piotr.stankiewicz@docker.com>
---------
Signed-off-by: Piotr Stankiewicz <piotr.stankiewicz@docker.com>
* added -o option to specify an output file name
* llama-tts returns ENOENT in case of file write error
note : PR #12042 is closed as superseded with this one.
* llama : introduce llama_set_warmup() API call that controls warmup mode; use all MoE experts during warmup
* common : use new API to enable warmup mode during model warmup
---------
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
As discussed in PR 'llama-tts : add -o option' (#12042):
* common_params : 'out_file' string is the only output file name parameter left in common_params. It's intended to be used in all example programs implementing an '-o' option.
* cvector-generator, export-lora, imatrix : default output filenames moved from 'common_params' to the 'main()' of each example program.
* sampler: turn lazy grammar trigger words to regexes
* add scripts/tool_bench.sh & .py
* constrain llama json output regardless of function name if matches at beginning
* update relaxed newline space rule in grammar tests
* support add_generation_prompt query parameter (useful for /apply_template)
* Update src/llama-grammar.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Add include files for std::min/max and std::toupper/tolower
* win32: move _USE_MATH_DEFINES before includes to ensure M_PI is defined
* Use GGML_RESTRICT instead of "restrict" keyword everywhere, and use "__restrict" in MSVC plain C mode
* win32: only use __restrict in MSVC if C11/C17 support is not enabled
---------
Co-authored-by: Marcus Groeber <Marcus.Groeber@cerence.com>
* Add chat template formatting to -no-cnv
* only enable prompt formatting if explicitly enabled
* add -st / --single-turn
* add --single-turn and -p in conversation mode
* fix -sys + -p
* reword warning
* small readability change and fix (long) outdated example usage
* only activate single turn in conversation mode
This commit adds a preset for llama.vim to use the default Qwen 2.5
Coder models.
The motivation for this change is to make it easier to start a server
suitable to be used with the llama.vim plugin. For example, the server
can be started with a command like the following:
```console
$ llama.vim --fim-qwen-1.5b-default
```
Refs: https://github.com/ggml-org/llama.cpp/issues/10932
* setup windows linking for llguidance; thanks @phil-scott-78
* add build instructions for windows and update script link
* change VS Community link from DE to EN
* whitespace fix
This commit adds completion for `--chat-template-file`, enabling only
`.jinja` files to be displayed as completions.
Example usage:
```console
$ ./build/bin/llama-cli --chat-template-file models/templates/<TAB>
models/templates/CohereForAI-c4ai-command-r7b-12-2024-tool_use.jinja
models/templates/CohereForAI-c4ai-command-r-plus-tool_use.jinja
models/templates/deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja
models/templates/deepseek-ai-DeepSeek-R1-Distill-Qwen-32B.jinja
models/templates/fireworks-ai-llama-3-firefunction-v2.jinja
models/templates/google-gemma-2-2b-it.jinja
models/templates/llama-cpp-deepseek-r1.jinja
models/templates/meetkai-functionary-medium-v3.1.jinja
models/templates/meetkai-functionary-medium-v3.2.jinja
models/templates/meta-llama-Llama-3.1-8B-Instruct.jinja
models/templates/meta-llama-Llama-3.2-3B-Instruct.jinja
models/templates/meta-llama-Llama-3.3-70B-Instruct.jinja
models/templates/microsoft-Phi-3.5-mini-instruct.jinja
models/templates/mistralai-Mistral-Nemo-Instruct-2407.jinja
models/templates/NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja
models/templates/NousResearch-Hermes-3-Llama-3.1-8B-tool_use.jinja
models/templates/Qwen-Qwen2.5-7B-Instruct.jinja
```
This is not limited to the models/templates directory, it can be used
anywhere in the filesystem, the above is just an example.
This commit adds a new option `--completion-bash` to the llama.cpp which
outputs a source-able bash completion script.
The motivation for this change is to provide a more user-friendly
experience for users who use the command-line interface of llama.cpp.
This is currently only basic and all options are displayed for all llama
executables but this can be improved in the future if needed.
Example usage:
```console
$ build/bin/llama-cli --completion-bash > ~/.llama-completion.bash
$ source ~/.llama-completion.bash
$ ./build/bin/llama-server --m<TAB>
--main-gpu --mirostat --mirostat-lr --model --multiline-input
--min-p --mirostat-ent --mlock --model-url
```
* extract & return thoughts in reasoning_content field (unless --reasoning-format) for DeepSeek R1 & Command R7B
* tool-calls: add deepseek r1 template (models/templates/llama-cpp-deepseek-r1.jinja) + hackommodate broken official template
* tool-calls: accommodate variety of wrong tool call opening tags both R1 Qwen 32B and 7B distills like to spit out
* server/oai: ensure content is null when there are tool calls, and reasoning_content appears before content for readability
* tool-calls: add DeepSeek R1 Qwen distills to server/README.md & server tests
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
The C API in llama.h claims users can implement `llama_sampler_i` to
create custom `llama_sampler`. The sampler chain takes ownership and
calls `llama_sampler_free` on them. However, `llama_sampler_free` is
hard-coded to use `delete`. This is undefined behavior if the object
wasn't also allocated via `new` from libllama's C++ runtime. Callers
in C and C-compatible languages do not use C++'s `new` operator. C++
callers may not be sharing the same heap as libllama.
* common : add default embeddings presets
This commit adds default embeddings presets for the following models:
- bge-small-en-v1.5
- e5-small-v2
- gte-small
These can be used with llama-embedding and llama-server.
For example, with llama-embedding:
```console
./build/bin/llama-embedding --embd-gte-small-default -p "Hello, how are you?"
```
And with llama-server:
```console
./build/bin/llama-server --embd-gte-small-default
```
And the embeddings endpoint can then be called with a POST request:
```console
curl --request POST \
--url http://localhost:8080/embeddings \
--header "Content-Type: application/json" \
--data '{"input": "Hello, how are you?"}'
```
I'm not sure if these are the most common embedding models but hopefully
this can be a good starting point for discussion and further
improvements.
Refs: https://github.com/ggerganov/llama.cpp/issues/10932
List devices in the same order as they appear when evaluating the model
and splitting tensors across devices, i.e. RPC devices come first in the
list.
ref #11435
Jonathan Graehl
Sigbjørn Skjæret
Copilot
Xuan-Son Nguyen
Sachin Desai
Georgi Gerganov
Diego Devesa
compilade
Jhen-Jie Hong
Aman Gupta
g2mt
kallewoof
Ed Addario
Molly Sophia
Eric Zhang
Alawode Oluwandabira
matteo
Ruikai Peng
aa956
fanyang
Piotr
Christian Kastner
bandoti
Olivier Chafik
Max Krasnyansky
Đinh Trọng Huy
Percy Piper
psocolovsky
Isaac McFadyen
Z
Olivier Chafik
Johannes Gäßler
David Huang
Helton Reis
Bartowski
Ycros
oobabooga
frob
Prajwal B Mehendarkar
yuri@FreeBSD
tastelikefeet
Sergey Fedorov
エシュナヴァリシア
R0CKSTAR
Michał Moskal
marcoStocchi
fairydreaming
mgroeber9110
dm4
Daniel Bevenius
standby24x7
Vinesh Janarthanan
JC
Maxim Evtush
Eric Curtin
Christian Fillion
Radoslav Gerganov