* Add Gemma3nVisionModel - MobileNetV5 vision encoder convertor to convert_hf_to_gguf.py. Add gemma3n to vision projectors in gguf-py/gguf/constants.py.
* Add mobilenetv5 impl
* Fix comments, remove unused vars
* Fix permute and remove transpose of projection weights
* Fix comments, remove debugging prints from hf_to_gguf
* 1. Hard-code image_mean = 0 and image_std = 1
2. Use available tensor mapping logic
3. Remove redundant chat template replacement of soft tokens placeholder with media placeholder
* 1. Move mobilenetv5 helpers declarations to `clip_graph_mobilenetv5` struct and definitions to mobilenetv5.cpp
2.Remove unused `clip_is_gemma3n` func declarations and definitions
3. Remove redundant `rescale_image_u8_to_f32` func and use `normalize_image_u8_to_f32` with zero mean and unit std
4. Calculate n_patches using image_size / patch_size
* Remove obsolete comments
* - convert_hf_to_gguf.py & constants.py & tensor_mapping.py: Use explicit mapping: Custom map for double indexed blocks and tensor_mapping.py for rest
- convert_hf_to_gguf.py: Unsqueeze Stem Bias and Layer scale tensors to correct shape while converting to gguf
- mobilenetv5.cpp: Remove explicit reshaping of Stem Bias and Layer scale which are now handled while converting to gguf, replace fprintf with LOG_*
- clip.cpp: Remove unused embedding and hard_emb_norm tensor loading
* - Rename tensors to v.conv..., v.blk..., v.msfa... to better align with already existing terminology
* Fix stem conv bias name
* Remove explicit handling of bias term for stem conv
* - Change order of addition in "project_per_layer_inputs" to support broadcasting of vision inp_per_layer
- Simplify the vision embeddings path of "get_per_layer_inputs" to output [n_embd_altup, n_layer, 1], broadcastable
* clean up conversion script
* fix code style
* also preserve audio tensors
* trailing space
* split arch A and V
* rm unused gemma3 func
* fix alignment
---------
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
* arg: support remote preset
* proof reading
* allow one HF repo to point to multiple HF repos
* docs: mention about multiple GGUF use case
* correct clean_file_name
* download: also return HTTP status code
* fix case with cache file used
* fix --offline option
* FlashAttention (#13)
* Add inplace softmax
* Move rms_norm to split row approach
* Update debug for supports_op
* clean up debug statements
* neg f16xf32xip builds and runs, havent actually ran a model that uses neg kernel yet though
* neg passes backend test
* unary operators pass ggml tests
* rms_norm double declaration bug atoned
* abides by editor-config
* removed vestigial files
* fixed autoconfig
* All operators (inlcluding xielu) working
* removed unnecesarry checking if node->src[1] exists for unary operators
* responded and dealt with PR comments
* implemented REPL_Template support and removed bug in unary operators kernel
* formatted embed wgsl and ggml-webgpu.cpp
* Faster tensors (#8)
Add fast matrix and matrix/vector multiplication.
* Use map for shader replacements instead of pair of strings
* Wasm (#9)
* webgpu : fix build on emscripten
* more debugging stuff
* test-backend-ops: force single thread on wasm
* fix single-thread case for init_tensor_uniform
* use jspi
* add pthread
* test: remember to set n_thread for cpu backend
* Add buffer label and enable dawn-specific toggles to turn off some checks
* Intermediate state
* Fast working f16/f32 vec4
* Working float fast mul mat
* Clean up naming of mul_mat to match logical model, start work on q mul_mat
* Setup for subgroup matrix mat mul
* Basic working subgroup matrix
* Working subgroup matrix tiling
* Handle weirder sg matrix sizes (but still % sg matrix size)
* Working start to gemv
* working f16 accumulation with shared memory staging
* Print out available subgroup matrix configurations
* Vectorize dst stores for sg matrix shader
* Gemv working scalar
* Minor set_rows optimization (#4)
* updated optimization, fixed errors
* non vectorized version now dispatches one thread per element
* Simplify
* Change logic for set_rows pipelines
---------
Co-authored-by: Neha Abbas <nehaabbas@macbookpro.lan>
Co-authored-by: Neha Abbas <nehaabbas@ReeseLevines-MacBook-Pro.local>
Co-authored-by: Reese Levine <reeselevine1@gmail.com>
* Comment on dawn toggles
* Working subgroup matrix code for (semi)generic sizes
* Remove some comments
* Cleanup code
* Update dawn version and move to portable subgroup size
* Try to fix new dawn release
* Update subgroup size comment
* Only check for subgroup matrix configs if they are supported
* Add toggles for subgroup matrix/f16 support on nvidia+vulkan
* Make row/col naming consistent
* Refactor shared memory loading
* Move sg matrix stores to correct file
* Working q4_0
* Formatting
* Work with emscripten builds
* Fix test-backend-ops emscripten for f16/quantized types
* Use emscripten memory64 to support get_memory
* Add build flags and try ci
---------
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
* Remove extra whitespace
* Move wasm single-thread logic out of test-backend-ops for cpu backend
* Disable multiple threads for emscripten single-thread builds in ggml_graph_plan
* Refactored pipelines and workgroup calculations (#10)
* refactored pipelines
* refactored workgroup calculation
* removed commented out block of prior maps
* Clean up ceiling division pattern
---------
Co-authored-by: Neha Abbas <nehaabbas@eduroam-169-233-141-223.ucsc.edu>
Co-authored-by: Reese Levine <reeselevine1@gmail.com>
* Start work on flash attention
* Shader structure set up (many bugs still)
* debugging
* Working first test
* Working with head grouping, head sizes to 128, logit softcap, mask/sinks enabled, f32
* Generalize softmax to work with multiple subgroups, f16 accumulation, mask shared memory tiling
* Start work on integrating pre-wgsl
* Separate structs/initial shader compilation library into separate files
* Work on compilation choices for flashattention
* Work on subgroup matrix/tile size portability
* subgroup size agnostic online softmax
* Cleanups, quantization types
* more cleanup
* fix wasm build
* Refactor flashattention to increase parallelism, use direct loads for KV in somce cases
* Checkpoint
* formatting
* Update to account for default kv cache padding
* formatting shader
* Add workflow for ggml-ci webgpu
* Try passing absolute path to dawn in ggml-ci
* Avoid error on device destruction, add todos for proper cleanup
* Fix unused warning
* Forgot one parameter unused
* Move some flashattn computation to f32 for correctness
* ggml: add env var GGML_OP_OFFLOAD_MIN_BATCH
* makes the min_batch_size for triggering op offload configurable via env var, defaulting to the prior hardcoded value of 32
* ggml: read GGML_OP_OFFLOAD_MIN_BATCH once and store to dev ctx
* cann: forward declaration of device context struct
* cann: move offload op check after device context declaration
* cuda: fix whitespace
Co-authored-by: Aman Gupta <amangupta052@gmail.com>
---------
Co-authored-by: Aman Gupta <amangupta052@gmail.com>
This commit adds a check comparing the installed transformers library
with the transformers version that the original model supports. This
check will be performed upon a model verification failure and prints a
warning/hint to the user suggesting to install the correct version of
the transformers library.
The motivation for this change is that it is possible for the model
verification to fail due to differences in the transformers library used
and it might not be obvious that this could be the cause of the failure.
With this warning the correct version can be checked and hopefully save
time troubleshooting the cause of the verification failure.
This commit removes the '-st` make target for running the converted
embedding model.
The motivation for this is that the pooling type is now part of the
.gguf metdata of the model and this is used by llama-debug when running
the model. So there is no need to specify the pooling type separately
any more.
The commit also adds an option to specify the type of normalization
applied to the output embeddings when running the converted model.
And the readme documentation has been updated to reflect these changes.
* Adding --direct-io flag for model loading
* Fixing read_raw() calls
* Fixing Windows read_raw_at
* Changing type off_t to size_t for windows and Renaming functions
* disable direct io when mmap is explicitly enabled
* Use read_raw_unsafe when upload_backend is available, not functional on some devices with Vulkan and SYCL
* Fallback to std::fread in case O_DIRECT fails due to bad address
* Windows: remove const keywords and unused functions
* Update src/llama-mmap.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: jtischbein <jtischbein@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Add libglvnd0, libgl1, libglx0, libegl1, libgles2 to the Vulkan
Dockerfile base image. These libraries are required by mesa-vulkan-drivers
to properly initialize the Vulkan ICD and detect GPU devices.
Without these libraries, vkEnumeratePhysicalDevices() returns an empty
list, resulting in "ggml_vulkan: No devices found." error.
Fixes#17761
* convert : clarify sentence-transformers-dense-modules help [no ci]
This commit updates this options help message which currently looks
like this:
```console
--sentence-transformers-dense-modules
Whether to include sentence-transformers dense modules.It can be used for sentence-transformers models, like
google/embeddinggemma-300mDefault these modules are not included.
```
* modify warptile tuning for xe3
* intel vendor check w/ coopmat support
* fix back formatting
* fix formatting change 2
* move intel check to chip specific tuning part
* Change to support both windows and linux
* modify m_warptile to l_warptile for intel
* modify warptile tuning for bf16 matmuls to fix regression (m_warptile to l_warptile)
* Code style changes
* Code style changes (2)
* Code style changes (3)
* examples : add debug utility/example
This commit introduces a new example named llama-debug which is a
utility that is intended to be used to assist with developing/debugging
a converted model.
The motivation for this utilitiy is to assist in model conversion work
to verify that the model produces the expected outputs. It is intended
to replace logits.cpp in examples/model-conversion.
Example usage:
```console
./build/bin/llama-debug \
-m models/Qwen2.5-0.5B-Instruct.gguf \
--prompt "Hello, my name is" \
--save-logits
...
Model add_bos: false
Input prompt: "Hello, my name is"
Token ids (5):
Hello(9707) ,(11) my(847) name(829) is(374)
Data saved to data/llamacpp-Qwen2.5-0.5B-Instruct.bin
Data saved to data/llamacpp-Qwen2.5-0.5B-Instruct.txt
Prompt saved to data/llamacpp-Qwen2.5-0.5B-Instruct-prompt.txt
Tokens saved to data/llamacpp-Qwen2.5-0.5B-Instruct-tokens.bin
```
For more details about the options available for this example, please
refer to examples/debug/README.md.
* throw runtime error instead of logging error
* remove params.warmup and enable the warmup/nowarmup option
* model-conversion : remove logits.cpp
This commit removes logits.cpp in favor of using llama-debug for
generating logits and embeddings.
* examples : remove model-conversion directory
This was missed in the previous commit.
* model-conversion : add support for saving prompt and token ids
This commit add support for storing the prompt and the token ids for the
prompt when running the original models.
The motivation for this is that this will allow us to compare the prompt
and the tokens generated for the prompt when verifing the converted
model. Currently it is possible that even if the same prompt is used
that the tokens generated are different if there is a difference in the
tokenization between the original and converted model which would
currently go unnoticed (the verification will most likely fail but it
might not be obvious why).
* squash! model-conversion : add support for saving prompt and token ids
fix pyright errors.
* model-conversion : add compare_tokens utility
This commit adds a script to compare token outputs between original and
converted models.
Example usage:
```console
(venv) $ ./scripts/utils/compare_tokens.py pytorch-gemma-3-270m-it llamacpp-gemma-3-270m-it-bf16
Comparing tokens between:
Original : pytorch-gemma-3-270m-it (6 tokens)
Converted: llamacpp-gemma-3-270m-it-bf16 (6 tokens)
✅ All 6 tokens match!
```
And there is a verbose flag that will also print out the prompts:
```console
(venv) $ ./scripts/utils/compare_tokens.py pytorch-gemma-3-270m-it llamacpp-gemma-3-270m-it-bf16 -v
Original model prompt (pytorch-gemma-3-270m-it):
prompt: Hello, my name is
n_tokens: 6
token ids: 2, 9259, 236764, 1041, 1463, 563
Converted model prompt (llamacpp-gemma-3-270m-it-bf16):
prompt: Hello, my name is
n_tokens: 6
token ids: 2, 9259, 236764, 1041, 1463, 563
Comparing tokens between:
Original : pytorch-gemma-3-270m-it (6 tokens)
Converted: llamacpp-gemma-3-270m-it-bf16 (6 tokens)
✅ All 6 tokens match!
```
* model-conversion : add token comparison to verifiction scripts
This commit add the calling of the compare_tokens function in
compare-logits.py and semantic_check.py to ensure that the token ids
that the tokenizers procoduce are the same before proceeding with
verifying the logits/embeddings.
Placing them in the existing scripts instead calling them separately
ensures that the token comparison is always done prior to the
logit/embedding verifications.
Follow up commit/pr could refactor the causal logits verification into
a single script instead of the two that exist now. This would reduce the
code and make it consistent with the embeddings verficiation which only
has a single script.
* debug : use llama_model_n_embd_out
This commit updates the debug example to use the new function
llama_model_n_embd_out instead of llama_model_n_embd.
The motivation for this change is to support late interation retriever
models, like LFM2-ColBert-350M, where the output embeddings are down
projected to a lower dimension.
* debug : add print_usage function
This commit adds a print_usage function that is passed to the
common_params_parse.
The motivation for this is that this enables a specific usage message
which will be printed after all the options, for example:
```console
example usage:
Print tensors:
./build/bin/llama-debug -m model.gguf -p "Hello my name is" --verbose
The tensors to be printed can be filtered with --tensor-filter option.
Save logits/embeddings:
./build/bin/llama-debug -m model.gguf -p "Hello my name is" --save-logits
Add --embedding to save embeddings
```
In #18624, get_env in ggml-cann was renamed to get_env_as_lowercase
to accurately reflect the function’s behavior and reduce the chance
of misuse. However, the update missed renaming call sites in other
files. This commit fixes that oversight.
* hexagon: improve fp16 matmul and add fp32/fp16 flash-attention
* hexagon: add support for set-rows fp32 -> fp16 with i32/i64 row-idx
* hexagon: add support for SCALE fp32
* hexagon: replace scalar fp32 -> fp16 copy with HVX
* hexagon: optimize flash_atten_ext with aligned VTCM buffers and DMA
- Implements double-buffered DMA prefetching for K, V, and Mask tensors.
- Ensures K and V rows in VTCM are padded to 128 bytes to support aligned HVX operations.
- Correctly synchronizes DMA transfers to prevent race conditions.
- Uses `FLASH_ATTN_BLOCK_SIZE` of 128 for efficient chunking.
* hexagon: use aligned mad_f16
* hexagon: flash_atten more aligned ops
* hexagon: optimize scale_f32 hvx helpers
* hexagon: unroll fa loops
* hexagon: remove unused set-rows log
* hexagon: flash_attn_ext add support for DMAing Q
- Update `op_flash_attn_ext` to include Q row size in scratchpad allocation.
- Pad Q row size to 128 bytes for alignment.
- Implement DMA transfer for Q tensor in `flash_attn_ext_f16_thread`.
- Update dot product computations to use VTCM-buffered Q data.
* hexagon: fix handling of NANs hvx dotproducts
* hexagon: cleanup spad allocation in flash-atten
* hexagon: improve fp16/fp32 matmul
- Introduced `vec_dot_f16_f16` and `vec_dot_f16_f16_rx2` kernels using efficient HVX dot product intrinsics.
- Added `quantize_fp32_f16` to copy/convert weights from DDR to VTCM
- Updated `op_matmul` to use the optimized path when VTCM capacity allows and broadcasting requirements are compatible.
- Implemented fallback logic to the original implementation for complex broadcasting scenarios.
* hexagon: fix HVX_ARCH check
* hexagon: matmul cleanup and fp16 fixes
Use aligned vec_dot_f16 for 2d matmuls and unaligned version for 4d.
* hexagon: fix fp16 x fp16 matmuls and some minor refactoring
* hexagon: add support for GET_ROWS f32 -> f32
Also optimize SET_ROWS threading a bit when we have just a few rows to process.
* hexagon: optimize set-rows threading
* hexagon: update adb/run-bench.sh to properly support experimental and verbose options
* hexagon: flash_atten use aligned vectors for dot products
Change is decoupled from https://github.com/ggml-org/llama.cpp/pull/18641.
[LFM2.5-Audio-1.5B](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B)
needs streaming istft for generating output audio.
* add streaming ISTFT class (`mtmd_audio_streaming_istft`) with overlap-add for audio reconstruction
* replace global audio cache with per-instance cache, the model requires
two independent caches, for preprocessing (audio input) and for istft
(audio output).
* unified templated FFT/IFFT implementation supporting both forward and inverse transforms
* vulkan: support buffer_from_host_ptr
* hacky use of buffer_from_host_ptr for directio
* disable buffer_from_host_ptr cap
* use external memory for ggml_vk_host_malloc, revert model loader changes
* disable external_memory_host for MoltenVK
* take buffer memory types into account
* don't use external_memory_host for ggml_vk_host_malloc
* server : add thinking content blocks to Anthropic Messages API
Add support for returning reasoning/thinking content in Anthropic API
responses when using models with --reasoning-format deepseek and the
thinking parameter enabled.
- Non-streaming: adds thinking block before text in content array
- Streaming: emits thinking_delta events with correct block indices
- Partial streaming: tracks reasoning state across chunks via
anthropic_has_reasoning member variable
Tested with bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF model.
* server : fix Anthropic API streaming for thinking content blocks
Add signature field and fix duplicate content_block_start events in
Anthropic Messages API streaming responses for reasoning models.
* server: refactor Anthropic streaming state to avoid raw pointer
Replace raw pointer to task_result_state with direct field copies:
- Copy state fields in update() before processing chunk
- Use local copies in to_json_anthropic() instead of dereferencing
- Pre-compute state updates for next chunk in update()
This makes the data flow clearer and avoids unsafe pointer patterns.
Xuan-Son Nguyen
Simranjeet Singh
shaofeiqi
Reese Levine
Pascal
Asbjørn Olling
Georgi Gerganov
Aaron Teo
Masashi Yoshimura
Jeff Bolz
Adrien Gallouët
도로로도로또
Johannes Gäßler
Doctor Shotgun
Daniel Bevenius
Julius Tischbein
Sigbjørn Skjæret
Oliver Walsh
R
virajwad
Eve
hipudding
Raul Torres
Max Krasnyansky
Tarek Dakhran
Aadeshveer Singh
Aman Gupta
Beinsezii
R
Christian Kastner