* ggml : improve ggml_mul speed when masking recurrent states
* llama : support running Mamba-Codestral-7B-v0.1
* llama : fix Mamba-2 conv state saving
* ggml : make the ggml_mul fast broadcast path more consistently formatted
* llama : remove unused variable
* llama : add missing break
* convert_hf : prefer SentencePiece tokenizer for Mamba-2 when present
The tokenzier.json of Mamba-Codestral-7B-v0.1 otherwise requires
workarounds to work correctly.
* llama : avoid redundant state copy for Mamba 1 and 2
* metal : attempt to adapt SSM_SCAN for Mamba-2
* metal : fix SSM_SCAN pipeline scope
* metal : use log and exp instead of log1pf and expf in SSM_SCAN
* metal : remove unused arguments for SSM_SCAN
The max index is 31, so trimming the arguments is necessary.
* metal : add back n_seqs to SSM_SCAN args
Whoops, this is needed for the offset in the concatenated output.
* metal : fix SSM_SCAN state head offset
* metal : fix wrong number of tokens per sequence in SSM_SCAN
* ggml : remove unused fast broadcast path in GGML_MUL
This was initially added because states were masked with ggml_mul,
but this is no longer done and so this "optimisation" is no longer
necessary, or at least not worth the additional code complexity.
* ggml : avoid multiply by D in GGML_OP_SSM_SCAN
This makes the weight buft detection in src/llama.cpp simpler.
* convert : transpose Mamba-2 A, D and reshape SSM_NORM
This breaks existing conversions of Mamba-2 models
to avoid some reshapes.
Not sure if it's a good idea,
but it makes the graph slightly cleaner.
* llama : more appropriate SSM_SCAN and SSM_CONV buft support checks
* convert : fix flake8 lint
* metal : fix confusion between ; and ,
* metal : add missing args for nb references in ssm_scan_f32_group
* metal : single-user mamba2 inference works
* kv-cache : remove const_cast when setting inputs for s_copy
And also fix multi-user inference for recurrent models
by using cell_id instead of i as the kv cell index
when populating s_copy.
* convert : avoid AutoConfig for Mamba and Mamba2 hparams
* kv-cache : allow context shift for recurrent models
* graph : fix recurrent state copies when avoiding copies
Works, but using lambda functions might not be that clean.
* ggml : fix mamba2 ssm scan when compiled with SVE
* ggml-cpu : reorder SVE FMA for consistency with other SIMD arches
* cuda : implement ssm scan for Mamba2
There is still room for improvement, but it works!
* cuda : adapt Mamba1 ssm scan to shape changes from Mamba2
* mamba : fix mismatched new and delete size for llm_build_mamba
Subclasses of llm_graph_context cannot have extra fields,
because the called destructor is not the one from the subclass.
This otherwise would cause problems when runnning Mamba-(1|2) inference
when compiled -DGGML_SANITIZE_ADDRESS=ON
* cuda : graceful fallback for Mamba-1 models with weird embd size
server : add dtw.params for v3-large-turbo (#3307)
* Add DTW model large-v3-turbo parameters to server.cpp example
DTW support is available in whispercpp and the large-v3-turbo model has already been added to the sources, but the large-v3-turbo model hasn't been added to the server.cpp file to make use of it. This commit hopefully corrects that issue.
* match original linebreak of original server.cpp file after adding large.v3.turbo dtw
* SYCL: Implement fused kernel GEGLU, SWIGLU and REGLU for single up+gate
* ggml : implement GLU for split up/gate (llama/14181)
* implement GLU for split up/gate
* add tests for ggml_glu_split
* Vulkan: Implement glu_split logic and shader support
* add split to logging [no ci]
* SYCL: refactor element_size ops and add split up and gate support to gated kernels
* SYCL: switch GEGLU to use tanh approximation
---------
Co-authored-by: 0cc4m <redacted> Co-authored-by: Akarshan <redacted>
* GGML: increase OP count in assertion
* Refactor: Optimize SYCL element-wise operations with unary function inlining
This commit refactors the SYCL element-wise operations to improve performance by:
- Inlining unary operations (sgn, abs, elu, gelu, silu, etc.) to reduce kernel launch overhead.
- Introducing helper functions `op_xxx` for each unary operation to encapsulate the logic.
- Replacing direct kernel calls with calls to these inlined functions.
- Using `__dpct_inline__` to encourage compiler inlining.
- Minor code cleanup and consistency improvements.
The changes aim to reduce kernel launch overhead and improve the overall efficiency of element-wise operations on SYCL devices.
* vulkan: Increase workgroup size for GLU, for performance (llama/14345)
* vulkan: Increase workgroup size for GLU, for performance
* vulkan: change GLU shaders to do one element per invocation rather than one row per workgroup
* merge fix
* metal : add support for split and swap
ggml-ci
---------
Co-authored-by: Georgi Gerganov <redacted> Co-authored-by: 0cc4m <redacted> Co-authored-by: Akarshan <redacted> Co-authored-by: Jeff Bolz <redacted>
Jeff Bolz [Sun, 29 Jun 2025 07:43:36 +0000 (02:43 -0500)]
vulkan: Add fusion support for RMS_NORM+MUL (llama/14366)
* vulkan: Add fusion support for RMS_NORM+MUL
- Add a use_count to ggml_tensor, so we can detect if an output is used more than once.
- Change the ggml-vulkan rms_norm shader to optionally multiply by another tensor.
- Add detection logic and basic fusion logic in ggml-vulkan.
- Add some testing support for fusion. Rather than computing one node at a time, allow
for computing the whole graph and just testing one node's results. Add rms_norm_mul tests
and enable a llama test.
* extract some common fusion logic
* fix -Winconsistent-missing-override
* move ggml_can_fuse to a common function
* build fix
* C and C++ versions of can_fuse
* move use count to the graph to avoid data races and double increments when used in multiple threads
* use hash table lookup to find node index
* change use_counts to be indexed by hash table slot
* minimize hash lookups
style fixes
* last node doesn't need single use.
fix type.
handle mul operands being swapped.
ggml-cpu : "align corners" for bilinear upscale/downscale (ggml/1285)
* add "align corners" mode for bilinear upscale, and allow downscaling
* add ggml_interpolate, deprecate ggml_upscale_ext, pass in align-corners as bit-flag
* test-backend-ops: replace ggml_upscale_ext with ggml_interpolate, add test cases for downscale and align-corners
Daniel Bevenius [Fri, 27 Jun 2025 13:43:56 +0000 (15:43 +0200)]
ci : use selective copy for musa image (#3296)
This commit modified the musa docker file to selectively copy
directories needed for the container image.
This commit also added a step to the docker workflow to free up disk
space in attempt to make enough room for the large musa build
containers.
The motivation for this change is to reduce the size of the container
image and try to avoid disk usage issues in CI.
Daniel Bevenius [Fri, 27 Jun 2025 07:55:56 +0000 (09:55 +0200)]
ci: set fail-fast to false in docker.yml (#3294)
* ci: set fail-fast to false in docker.yml
This commit modifies the GitHub Actions workflow for Docker builds to
disable the fail-fast behavior.
The motivation for this is that currently if one of the strategy jobs
fails any other job that is in progress will be cancelled. There is no
need for this as the jobs are independent.
* ci : update docker.yml to use a single build
This commit updates the docker job to only build the image once instead
of twice (only happens when pushing to the master branch). Instead this
will tag the image with the commit SHA when pushing to master.
The motivation for this change is to reduce the time it takes to run
this job and also it might help with the disk space issues we are
experiencing for this job when it runs on pushes to master.
Daniel Bevenius [Thu, 26 Jun 2025 14:29:29 +0000 (16:29 +0200)]
ci : add should_release variable (#3288)
* ci : add should_release variable
This commit adds a `should_release` variable to the GitHub Actions
workflow to determine if a release should be created based on the tag or
branch conditions.
The motivation for this that it simplifies the logic for deciding
whether to upload artifacts or not, making it easier to maintain if we
need to change the conditions in the future.
Daniel Bevenius [Wed, 25 Jun 2025 19:43:58 +0000 (21:43 +0200)]
ci : add support for tag-based releases (#3287)
This commit modifies the GitHub Actions workflow to support
tag-based releases. When a tag is pushed that starts with 'v', the
workflow will use that tag name for the release process.
I think this was the once the behavior, but it was lost in updates that
I've made to the workflow. This commit restores that functionality.
Daniel Bevenius [Wed, 25 Jun 2025 12:16:31 +0000 (14:16 +0200)]
stream : add nullptr check of whisper_context (#3283)
* stream : add nullptr check of whisper_context
This commit adds a check to ensure that the `whisper_context` is not
null after initialization.
The motivation for this is that currently, if the initialization fails,
the program continues to run leading to a segmentation fault. This sort
of check is performed by others examples like whisper-cli.
Daniel Bevenius [Wed, 25 Jun 2025 10:12:36 +0000 (12:12 +0200)]
ci : enable main-cuda build (#3282)
This commit re-enables the main-cuda Docker build in the CI workflow.
The main-cuda Dockerfile has been updated to remove build artifacts
and also print the size of the /app directory after the build. A similar
change was recently made to the musa Dockerfile, and perhaps this job
was also having similar disk space issues.
The motivation for this change is that this configuration has been
disabled for a while due to persistent build failures. However, the
actual logs are now longer available.
Daniel Bevenius [Tue, 24 Jun 2025 06:20:28 +0000 (08:20 +0200)]
ci : reduce musa image size (#3277)
* ci : reduce musa image size
This commit contains an attempt to reduce the size of the musa Docker
image by copying only the necessary files from the build stage.
The motivation for this is that the CI runs sometimes fail with out of
memory errors. These seems to be able to pass for PRs, at least
sometimes but fail upon push to the master branch.
* ci : remove build time files instead of selective copying
Daniel Bevenius [Mon, 23 Jun 2025 10:34:44 +0000 (12:34 +0200)]
ci : add apt-get clean to musa Dockerfile (#3275)
* ci : add apt-get clean to musa Dockerfile
This commit adds `apt-get clean` to the musa Dockerfile to reduce the
image size by removing cached package files after installation.
The motivation for this is to try to reduce the size of the Docker image
and see if this can avoid the "no space left on device" error during
the CI build process.
Nicolò Scipione [Fri, 20 Jun 2025 13:07:21 +0000 (15:07 +0200)]
sycl: add usage of enqueue_functions extension (llama/14244)
* Add header and namespace to use enqueue_functions extension
* Convert submit and parallel_for to use new extension in convert.cpp
* Convert submit and parallel_for to use extension in ggml-sycl.cpp
* Convert submit and parallel_for to use extension in gla.cpp
* Convert submit and parallel_for in mmq.cpp
* Convert submit and parallel_for in mmvq.cpp
* Convert submit and parallel_for in remaining files
* Convert all simple parallel_for to nd_launch from enqueue_functions
extension
* Wrapping extension in general function
Create a general function that enable the enqueue_functions extension if
it is enable in the compiler, otherwise call the general SYCL function
to launch kernels.
* android.java : update cmake to use FetchContent for ggml
This commit updates the CMake configuration for the Android Java example
to use `FetchContent` for including the `ggml` library. Do be able to
use FetchContent we also update the `compileSdkVersion` and
`targetSdkVersion` to 31, and the `buildToolsVersion` to '30.0.3'.
This also required a an update to the Gradle plugin version to 7.4.0.
The motivation for this change is avoid having to make manual changes to
the CMakeLists.txt file after syncing the ggml library.
Daniel Bevenius [Wed, 18 Jun 2025 15:41:43 +0000 (17:41 +0200)]
examples : add stereo to mono conversion in read_audio_data (#3266)
This commit adds a conversion from stereo to mono in the
`read_audio_data` function of `common-whisper.cpp`.
The motivation for this change is prior to Commit 7d3da68f792018e81a758881e081154d1cbe6b6f ("examples : use miniaudio for
direct decoding flac, mp3, ogg and wav (#2759)", there was a step that
read stereo int16 data -> pcm16 (448512 samples), and then converted to
mono (224256 samples), and then also convert to stereo in `pcmf32s.
The middle step here seems to have been missed when rewriting the code to
use Miniaudio and caused issues then transcribing stereo audio files.
For example, currently using the audio sample in the linked issue the
output is:
```console
[00:00:00.000 --> 00:00:03.000] (speaker 1) Sous-titres réalisés para la communauté d'Amara.org
```
And with the change in this commit the output is:
```
[00:00:00.000 --> 00:00:01.500] (speaker 1) *sonnerie de téléphone*
[00:00:01.500 --> 00:00:07.000] (speaker 1) Salut jeune homme !
[00:00:07.000 --> 00:00:08.500] (speaker 0) C'est vrai que je te dérange ?
[00:00:08.500 --> 00:00:10.500] (speaker 1) Ah pas du tout, pas du tout, pas du tout !
[00:00:10.500 --> 00:00:12.500] (speaker 1) J'étais en train de...
[00:00:12.500 --> 00:00:14.500] (speaker 1) de préparer un courrier
```
overview / pkg / ggml / sources / whisper.cpp / log