]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/log
pkg/ggml/sources/llama.cpp
2 months agosync : ggml
Georgi Gerganov [Thu, 10 Apr 2025 20:59:16 +0000 (23:59 +0300)]
sync : ggml

ggml-ci

2 months agoggml: don't include arm_neon.h when using CUDA 12 with ARM Neon (ggml/1187)
cmdr2 [Thu, 10 Apr 2025 12:23:08 +0000 (17:53 +0530)]
ggml: don't include arm_neon.h when using CUDA 12 with ARM Neon (ggml/1187)

fix #1186

2 months agoggml : add bilinear upscale support (ggml/1185)
Diego Devesa [Wed, 9 Apr 2025 10:32:13 +0000 (12:32 +0200)]
ggml : add bilinear upscale support (ggml/1185)

2 months agoggml : add more generic custom op, remove deprecated custom ops (ggml/1183)
Diego Devesa [Wed, 9 Apr 2025 10:31:34 +0000 (12:31 +0200)]
ggml : add more generic custom op, remove deprecated custom ops (ggml/1183)

* ggml : add more generic ggml_custom op

* ggml : remove deprecated custom ops

2 months agoscripts : fix sync-ggml-am.sh
Georgi Gerganov [Thu, 10 Apr 2025 20:59:01 +0000 (23:59 +0300)]
scripts : fix sync-ggml-am.sh

2 months agollava : introduce libmtmd (#12849)
Xuan-Son Nguyen [Thu, 10 Apr 2025 20:57:16 +0000 (22:57 +0200)]
llava : introduce libmtmd (#12849)

* wip llava2

* migrated gemma3 to llava2

* add timings

* correct pre/postfix

* fix missing include

* fix compilation unused var warn

* update llava2_tokenize

* change name llava2 --> mtmd

* improve api

* refine helpers

* Update examples/llava/mtmd.cpp

Co-authored-by: Georgi Gerganov <redacted>
---------

Co-authored-by: Georgi Gerganov <redacted>
2 months agoconvert : ability to lazy-load safetensors remotely without downloading to disk ...
Xuan-Son Nguyen [Thu, 10 Apr 2025 15:24:44 +0000 (17:24 +0200)]
convert : ability to lazy-load safetensors remotely without downloading to disk (#12820)

* gguf util : add SafetensorRemote

* fix style

* convert: add --remote option

* convert : allow using lazy remote tensors

It's a bit slow for now since everything is blocking and single-threaded.

* correct metadata.name

* small style fix

* support HF_TOKEN

* convert : use writeable buffer for remote lazy tensors

* convert : fix flake8 lint regarding lamdba assigment

* multithreaded download

* multithread: print debug

* fix style

* Revert "multithreaded download"

This reverts commit 42fc895ace385edc972ad819c76c704aeea61791.

* bring back _get_request_headers

---------

Co-authored-by: Francis Couture-Harpin <redacted>
2 months agoCANN: Support more ops (#12841)
Chenguang Li [Thu, 10 Apr 2025 00:51:52 +0000 (08:51 +0800)]
CANN: Support more ops (#12841)

* [CANN]Support Opt LOG && MEAN && PAD_REFLECT_1D

* [CANN]Support COUNT_EQUAL && STEP && SGN

* [CANN]codestyle adjustment

* [CANN]codestyle adjustment

---------

Signed-off-by: noemotiovon <redacted>
2 months agoFixes #12823 (#12830)
Prajwal B Mehendarkar [Wed, 9 Apr 2025 23:18:01 +0000 (04:48 +0530)]
Fixes #12823 (#12830)

* Including limits file on AIX

* Fixes #12823

2 months agodocker : added all CPU to GPU images (#12749)
Rudi Servo [Wed, 9 Apr 2025 23:17:12 +0000 (23:17 +0000)]
docker : added all CPU to GPU images (#12749)

2 months agoggml-cpu-impl.h: do not redefine bool on POWER9 (#12856)
Piotr Kubaj [Wed, 9 Apr 2025 23:00:34 +0000 (23:00 +0000)]
ggml-cpu-impl.h: do not redefine bool on POWER9 (#12856)

error: unknown type name '_Bool'

2 months agoggml-impl.h: fix build on POWER9 (#12855)
Piotr Kubaj [Wed, 9 Apr 2025 23:00:25 +0000 (23:00 +0000)]
ggml-impl.h: fix build on POWER9 (#12855)

error: ISO C++17 does not allow 'register' storage class specifier

2 months agollama : Support Qwen3 and Qwen3MoE (#12828)
Bo Zheng [Wed, 9 Apr 2025 09:47:36 +0000 (17:47 +0800)]
llama : Support Qwen3 and Qwen3MoE (#12828)

* add qwen3 & qwen3moe support.

* fix

---------

Co-authored-by: bozheng-hit <redacted>
2 months agomusa: enable freediskspace for docker image build (#12839)
R0CKSTAR [Wed, 9 Apr 2025 09:22:30 +0000 (17:22 +0800)]
musa: enable freediskspace for docker image build (#12839)

Signed-off-by: Xiaodong Ye <redacted>
2 months agosycl: update documentation to use -no-cnv (#12845)
Romain Biessy [Wed, 9 Apr 2025 09:22:04 +0000 (11:22 +0200)]
sycl: update documentation to use -no-cnv (#12845)

2 months agoci: detach common from the library (#12827)
Plamen Minev [Wed, 9 Apr 2025 08:11:11 +0000 (11:11 +0300)]
ci: detach common from the library (#12827)

* fix: detach common from the library

* fix: building chat test template

2 months agoclip : do not print ftype (#12832)
Xuan-Son Nguyen [Wed, 9 Apr 2025 08:09:53 +0000 (10:09 +0200)]
clip : do not print ftype (#12832)

2 months agoreadme : add rpc backend (#12842)
Georgi Gerganov [Wed, 9 Apr 2025 07:54:42 +0000 (10:54 +0300)]
readme : add rpc backend (#12842)

2 months agoCANN: Support Opt CONV_TRANSPOSE_1D and ELU (#12786)
Chenguang Li [Wed, 9 Apr 2025 06:04:14 +0000 (14:04 +0800)]
CANN: Support Opt CONV_TRANSPOSE_1D and ELU (#12786)

* [CANN] Support ELU and CONV_TRANSPOSE_1D

* [CANN]Modification review comments

* [CANN]Modification review comments

* [CANN]name adjustment

* [CANN]remove lambda used in template

* [CANN]Use std::func instead of template

* [CANN]Modify the code according to the review comments

---------

Signed-off-by: noemotiovon <redacted>
2 months agovulkan: In coopmat2 mmq, load q4_k/q5_k scales through shared memory (#12833)
Jeff Bolz [Wed, 9 Apr 2025 05:25:08 +0000 (00:25 -0500)]
vulkan: In coopmat2 mmq, load q4_k/q5_k scales through shared memory (#12833)

q4_k and q5_k had a lot of redundant global loads where the same 16B of
scale information is repeatedly loaded and decoded during each loop iteration.
This change restructures the loops to more explicitly iterate over whole
blocks in the outer loop (with unrolled inner loop) and to copy/decode the
scale data into shared memory once at the start of each outer loop. The copy
is pipelined so the scale load from global memory is relatively cheap.

This improves q4_k/q5_k model prompt processing performance by around 5-7%.
I briefly tried applying this to q6_k and q4_0, and it didn't help for q6_k
and hurt for q4_0.

The big "else" path in mul_mm_cm2.comp that had all the clamped/unclamped
variants isn't used as often as it originally was (e.g. due to the padded_N
change), so I trimmed it down to offset some of the new complexity of the
semi-manual loop unrolling.

2 months agovulkan: Use fp16 for the flash attention P*V multiplication (#12783)
Jeff Bolz [Wed, 9 Apr 2025 05:12:57 +0000 (00:12 -0500)]
vulkan: Use fp16 for the flash attention P*V multiplication (#12783)

This is consistent with the ggml-cuda behavior and the mul_mat fallback.

2 months agocuda : add f32 to bf16 copy op (#12806)
Sigbjørn Skjæret [Tue, 8 Apr 2025 21:21:31 +0000 (23:21 +0200)]
cuda : add f32 to bf16 copy op (#12806)

This allows BF16 KV-cache on CUDA.

2 months agollava: improve clip_ctx destructor to not memleak load_image_size (#12834)
Matt Clayton [Tue, 8 Apr 2025 20:01:58 +0000 (16:01 -0400)]
llava: improve clip_ctx destructor to not memleak load_image_size (#12834)

2 months agollama : fix FA when KV cache is not used (i.e. embeddings) (#12825)
Georgi Gerganov [Tue, 8 Apr 2025 16:54:51 +0000 (19:54 +0300)]
llama : fix FA when KV cache is not used (i.e. embeddings) (#12825)

* ggml : FA supports F32 V

* graph : cast KV to F16 when the KV cache is not used

ggml-ci

* server : add test that exercises embeddings with FA enabled

ggml-ci

2 months agoserver : fix thread.join() on exit (#12831)
Xuan-Son Nguyen [Tue, 8 Apr 2025 16:37:06 +0000 (18:37 +0200)]
server : fix thread.join() on exit (#12831)

2 months agollava: add more helper functions to check projector types in clip context (#12824)
dm4 [Tue, 8 Apr 2025 13:49:13 +0000 (21:49 +0800)]
llava: add more helper functions to check projector types in clip context (#12824)

Signed-off-by: dm4 <redacted>
2 months agoarg : Including limits file on AIX (#12822)
Prajwal B Mehendarkar [Tue, 8 Apr 2025 12:30:59 +0000 (18:00 +0530)]
arg : Including limits file on AIX (#12822)

2 months agoserver : webui : Improve Chat Input with Auto-Sizing Textarea (#12785)
characharm [Tue, 8 Apr 2025 09:14:59 +0000 (14:14 +0500)]
server : webui : Improve Chat Input with Auto-Sizing Textarea (#12785)

* Update ChatScreen.tsx

* useAutosizeTextarea.ts

useAutosizeTextarea to encapsulate the logic.

* Implement responsive auto-sizing chat textarea

Replaces the manual textarea resizing with an automatic height adjustment based on content.

- `useChatTextarea` hook to manage textarea state and auto-sizing logic via refs, preserving the optimization
- Textarea now grows vertically up to a maximum height (`lg:max-h-48`) on large screens (lg breakpoint and up).
- Disables auto-sizing and enables manual vertical resizing (`resize-vertical`) on smaller screens for better mobile usability.
- Aligns the "Send" button to the bottom of the textarea (`items-end`) for consistent positioning during resize.

* -update compressed index.html.gz after npm run build
-refactor: replace OptimizedTextareaValue with AutosizeTextareaApi in VSCode context hook

* chore: normalize line endings to LF
refactor: AutosizeTextareaApi -> chatTextareaApi

* refactor: Rename interface to PascalCase

---------

Co-authored-by: Xuan Son Nguyen <redacted>
2 months agoRevert "sycl:remove redundant memcopy in function ggml_backend_sycl_buffer_set_tensor...
Neo Zhang Jianyu [Tue, 8 Apr 2025 07:03:21 +0000 (15:03 +0800)]
Revert "sycl:remove redundant memcopy in function ggml_backend_sycl_buffer_set_tensor" (#12812)

* Revert "sycl: remove redundant memcopy in function ggml_backend_sycl_buffer_s…"

This reverts commit 518a01480eb3a7c80a4951b430db9dee55428310.

* Update ggml/src/ggml-sycl/ggml-sycl.cpp

* Update ggml/src/ggml-sycl/ggml-sycl.cpp

* rm tail space

2 months agogguf-py : support lazy tensor splitting (#12809)
compilade [Tue, 8 Apr 2025 07:03:07 +0000 (03:03 -0400)]
gguf-py : support lazy tensor splitting (#12809)

* gguf-py : support lazy tensor splitting

Splitting usually involves returning tuples of tensors,
which need to be handled properly to avoid early eager evaluation.

* gguf-py : fix flake8 lint

2 months agollama : Support llama 4 text-only (#12791)
Xuan-Son Nguyen [Mon, 7 Apr 2025 21:06:44 +0000 (23:06 +0200)]
llama : Support llama 4 text-only (#12791)

* llama4 conversion

* initial support, no chat template

* clean up a bit

* fix tokenizer conversion

* correct hparams

* try this

* fix shexp

* ffn_inp_normed

* chat template

* clean up model conversion

* add_bos

* add scale_before_ffn

* fix order

* weight_before_ffn

* llm_graph_input_attn_temp

* add chunk attn mask

* build_inp_attn_scale()

* add comment about ggml_repeat

* clarify comments

* fix build

2 months agoopencl: better identify Adreno GPU (#12760)
lhez [Mon, 7 Apr 2025 20:22:54 +0000 (13:22 -0700)]
opencl: better identify Adreno GPU (#12760)

2 months agohellaswag: display estimated score confidence interval (#12797)
stduhpf [Mon, 7 Apr 2025 15:47:08 +0000 (17:47 +0200)]
hellaswag: display estimated score confidence interval (#12797)

2 months agocuda : fix HIP and MUSA BF16 (#0)
Georgi Gerganov [Mon, 7 Apr 2025 10:18:07 +0000 (13:18 +0300)]
cuda : fix HIP and MUSA BF16 (#0)

ggml-ci

2 months agosync : ggml
Georgi Gerganov [Mon, 7 Apr 2025 09:32:39 +0000 (12:32 +0300)]
sync : ggml

ggml-ci

2 months agoggml : simplify Arm fp16 CPU logic (ggml/1177)
Georgi Gerganov [Mon, 7 Apr 2025 09:25:15 +0000 (12:25 +0300)]
ggml : simplify Arm fp16 CPU logic (ggml/1177)

* ggml : simlpify Arm fp16 CPU logic

ggml-ci

* cont : bring back CUDA/MUSA checks

ggml-ci

2 months agoCUDA: don't convert BF16 weights to FP32 (ggml/1174)
Sigbjørn Skjæret [Fri, 4 Apr 2025 19:05:12 +0000 (21:05 +0200)]
CUDA: don't convert BF16 weights to FP32 (ggml/1174)

* add bf16 support

* use convert_from_bf16_cuda instead of convert_unary_cuda for f32

* revert 7ec5085

* move functionality into convert_unary with constexpr

2 months agocpu: move all the operators into a separate c++ file (except mul_mat) (ggml/1167)
cmdr2 [Wed, 2 Apr 2025 12:16:16 +0000 (17:46 +0530)]
cpu: move all the operators into a separate c++ file (except mul_mat) (ggml/1167)

* cpu: refactor SIMD mappings and vectorized op functions into separate files

* Fix warning for ggml_float to float

* Fix warnings

* cpu: move all the operations (except mul_mat) to a separate c++ file

* fix whitespace

* Update ggml/src/ggml-cpu/vec.h

Co-authored-by: Diego Devesa <redacted>
* Fix PR comments - use GGML_UNUSED, use cassert in ops.cpp

* Reverse the order of import for ops.h and vec.h, to match what was present in ggml-cpu.c previously

---------

Co-authored-by: Diego Devesa <redacted>
2 months agosycl: remove redundant memcopy in function ggml_backend_sycl_buffer_set_tensor (...
zhouwg [Mon, 7 Apr 2025 15:22:57 +0000 (23:22 +0800)]
sycl: remove redundant memcopy in function ggml_backend_sycl_buffer_set_tensor (#12734)

2 months agoci : no curl on ggml-ci (#12796)
Xuan-Son Nguyen [Mon, 7 Apr 2025 12:37:28 +0000 (14:37 +0200)]
ci : no curl on ggml-ci (#12796)

2 months agocmake : enable curl by default (#12761)
Xuan-Son Nguyen [Mon, 7 Apr 2025 11:35:19 +0000 (13:35 +0200)]
cmake : enable curl by default (#12761)

* cmake : enable curl by default

* no curl if no examples

* fix build

* fix build-linux-cross

* add windows-setup-curl

* fix

* shell

* fix path

* fix windows-latest-cmake*

* run: include_directories

* LLAMA_RUN_EXTRA_LIBS

* sycl: no llama_curl

* no test-arg-parser on windows

* clarification

* try riscv64 / arm64

* windows: include libcurl inside release binary

* add msg

* fix mac / ios / android build

* will this fix xcode?

* try clearing the cache

* add bunch of licenses

* revert clear cache

* fix xcode

* fix xcode (2)

* fix typo

2 months agoCANN: fix typo in ggml-cann (#12733)
zhouwg [Mon, 7 Apr 2025 11:34:14 +0000 (19:34 +0800)]
CANN: fix typo in ggml-cann (#12733)

2 months agoCANN: Refactor to reduce duplicate code (#12731)
hipudding [Mon, 7 Apr 2025 09:10:36 +0000 (17:10 +0800)]
CANN: Refactor to reduce duplicate code (#12731)

* CANN: Refactor to reduce duplicate code

* CANN: fix review comment

2 months agomusa: fix compilation warnings in mp_22/31 (#12780)
R0CKSTAR [Sun, 6 Apr 2025 13:23:54 +0000 (21:23 +0800)]
musa: fix compilation warnings in mp_22/31 (#12780)

Signed-off-by: Xiaodong Ye <redacted>
2 months agovulkan: fix NaN issue in flash attention shader (#12776)
Jeff Bolz [Sun, 6 Apr 2025 09:03:47 +0000 (04:03 -0500)]
vulkan: fix NaN issue in flash attention shader (#12776)

Use -FLT_MAX/2 rather than -inf as the initial value for computing the maximum.

2 months agovulkan: Use unclamped loads for flash attention mask (#12720)
Jeff Bolz [Sun, 6 Apr 2025 08:47:13 +0000 (03:47 -0500)]
vulkan: Use unclamped loads for flash attention mask (#12720)

nem1 must be a multiple of GGML_KQ_MASK_PAD, and GGML_KQ_MASK_PAD is a multiple
of the number of rows in the matrix. The KV dim is a multiple of the number of
columns for the aligned shader.

2 months agoVulkan: Tune Vulkan mmq int dot shader for performance (#12767)
0cc4m [Sat, 5 Apr 2025 16:04:03 +0000 (18:04 +0200)]
Vulkan: Tune Vulkan mmq int dot shader for performance (#12767)

2 months agocommon : fix includes in arg.cpp and gemma3-cli.cpp (#12766)
Sergey Fedorov [Sat, 5 Apr 2025 15:46:00 +0000 (23:46 +0800)]
common : fix includes in arg.cpp and gemma3-cli.cpp (#12766)

* arg.cpp: add a missing include

* gemma3-cli.cpp: fix cinttypes include

2 months agoclip : refactor clip_init, add tests (#12757)
Xuan-Son Nguyen [Sat, 5 Apr 2025 15:17:40 +0000 (17:17 +0200)]
clip : refactor clip_init, add tests (#12757)

* refactor clip_init

* fix loading file

* fix style

* test ok

* better test with report

* add missing headers

* clarify

* add KEY_MM_PATCH_MERGE_TYPE

* remove bool has_* pattern

* Apply suggestions from code review

Co-authored-by: Georgi Gerganov <redacted>
* Update examples/llava/clip.cpp

Co-authored-by: Georgi Gerganov <redacted>
* use ggml_soft_max_ext

* refactor logging system

* add minicpm-v-o 2.6 for testing

* use nullptr everywhere

* fix Yi-VL model

---------

Co-authored-by: Georgi Gerganov <redacted>
2 months agocommon: custom hf endpoint support (#12769)
エシュナヴァリシア [Sat, 5 Apr 2025 13:31:42 +0000 (21:31 +0800)]
common: custom hf endpoint support (#12769)

* 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 <redacted>
* Apply suggestions from code review

---------

Co-authored-by: ベアトリーチェ <redacted>
Co-authored-by: Xuan-Son Nguyen <redacted>
2 months agosync: minja (#12739)
Olivier Chafik [Fri, 4 Apr 2025 20:16:39 +0000 (13:16 -0700)]
sync: minja (#12739)

* sync: minja

https://github.com/google/minja/pull/57

* fix json include

2 months agokv-cache : simplify + fix warning for recurrent models (#12756)
Georgi Gerganov [Fri, 4 Apr 2025 18:48:10 +0000 (21:48 +0300)]
kv-cache : simplify + fix warning for recurrent models (#12756)

ggml-ci

2 months agoci: add Linux cross-compile build (#12428)
bandoti [Fri, 4 Apr 2025 17:05:12 +0000 (14:05 -0300)]
ci: add Linux cross-compile build (#12428)

2 months agoserver : webui : Upgrade daisyui, tailwindcss. (#12735)
Nauful Shaikh [Fri, 4 Apr 2025 14:09:52 +0000 (09:09 -0500)]
server : webui : Upgrade daisyui, tailwindcss. (#12735)

* Upgrade daisyui, tailwindcss.

* Switch to all themes.

* Revert a change.

* Update formatting.

* Install packages before npm build.

* Revert "Install packages before npm build."

This reverts commit 336c5147e614e60993162794ba9d9d4629a916f8.

* Add index.html.gz

* run build

---------

Co-authored-by: Xuan Son Nguyen <redacted>
2 months agogguf-split : --merge now respects --dry-run option (#12681)
nick huang [Fri, 4 Apr 2025 14:09:12 +0000 (22:09 +0800)]
gguf-split : --merge now respects --dry-run option (#12681)

* gguf-split now respects dry-run option

* removing trailing space

2 months agosycl: allow ggml-sycl configuration and compilation using Visual Studio project/solut...
Nicolò Scipione [Fri, 4 Apr 2025 14:00:46 +0000 (16:00 +0200)]
sycl: allow ggml-sycl configuration and compilation using Visual Studio project/solution (#12625)

2 months agocmake: fix ggml-shaders-gen compiler paths containing spaces (#12747)
Ronny Brendel [Fri, 4 Apr 2025 13:12:40 +0000 (15:12 +0200)]
cmake: fix ggml-shaders-gen compiler paths containing spaces (#12747)

fixes error for compiler paths with spaces

2 months agodocs : add XCFramework section to README.md [no ci] (#12746)
Daniel Bevenius [Fri, 4 Apr 2025 08:24:12 +0000 (10:24 +0200)]
docs : add XCFramework section to README.md [no ci] (#12746)

This commit adds a new section to the README.md file, detailing the
usage of the XCFramework.

The motivation for this is that it might not be immediately clear to
users how to use the XCFramework in their projects and hopefully this
will help.

2 months agovulkan: Hybrid waitForFences/getFenceStatus to reduce fence latency (#12630)
Jeff Bolz [Fri, 4 Apr 2025 05:54:35 +0000 (00:54 -0500)]
vulkan: Hybrid waitForFences/getFenceStatus to reduce fence latency (#12630)

There seems to be a bubble waking up from waitForFences, which costs a few
percent performance and also increased variance in performance. This change
inserts an "almost_ready" fence when the graph is about 80% complete and we
waitForFences for the almost_ready fence and then spin (with _mm_pauses) waiting
for the final fence to be signaled.

2 months agovulkan: set cmake minimum and project name in vulkan-shaders (#12744)
Jeff Bolz [Fri, 4 Apr 2025 05:53:20 +0000 (00:53 -0500)]
vulkan: set cmake minimum and project name in vulkan-shaders (#12744)

2 months agoopencl: update doc for OpenCL (#12702)
lhez [Fri, 4 Apr 2025 05:18:17 +0000 (22:18 -0700)]
opencl: update doc for OpenCL (#12702)

* opencl: add OpenCL to build.md

* opencl: remove fixed issue/TODO

* opencl: add link to OPENCL.md

* opencl: update doc - refine tools requirement for Windows 11 arm64

2 months agoCUDA: Prefer vector flash decoding kernel for Gemma models (#12738)
Gaurav Garg [Thu, 3 Apr 2025 16:20:29 +0000 (21:50 +0530)]
CUDA: Prefer vector flash decoding kernel for Gemma models (#12738)

* Prefer vector flash decoding kernel for Gemma models

Vector flash decoding kernel was not being picked for models with head dimension 256. Gemma models are in this category.
Removing this limit improves e2e performance by upto 12% in gen phase throughput for Gemm models.

* Update ggml/src/ggml-cuda/fattn.cu

Co-authored-by: Johannes Gäßler <redacted>
---------

Co-authored-by: Johannes Gäßler <redacted>
2 months agovocab : use string_view::find() to avoid unnecessary looking up beyond the fragment...
yumeyao [Thu, 3 Apr 2025 15:32:54 +0000 (23:32 +0800)]
vocab : use string_view::find() to avoid unnecessary looking up beyond the fragment range (#12706)

2 months agovulkan: Fix missing cmake logic for dot product extension (#12721)
Jeff Bolz [Thu, 3 Apr 2025 15:08:26 +0000 (10:08 -0500)]
vulkan: Fix missing cmake logic for dot product extension (#12721)

2 months agoci : add env variable in ggml-ci and document the same in SYCL.md (#12736)
Atharva Dubey [Thu, 3 Apr 2025 12:12:39 +0000 (13:12 +0100)]
ci : add env variable in ggml-ci and document the same in SYCL.md (#12736)

2 months agosync : minja (inclusionAI/Ling) and update tests (#12699)
R0CKSTAR [Thu, 3 Apr 2025 11:51:35 +0000 (19:51 +0800)]
sync : minja (inclusionAI/Ling) and update tests (#12699)

Signed-off-by: Xiaodong Ye <redacted>
2 months agofix MUSA compiler warning (#12704)
a3sh [Thu, 3 Apr 2025 07:32:55 +0000 (15:32 +0800)]
fix MUSA compiler warning (#12704)

* fix MUSA compiler warning

* replace (void) with GGML_UNUSED

2 months agoCANN: Support operator SIN COS ARGMAX (#12709)
Chenguang Li [Thu, 3 Apr 2025 07:18:08 +0000 (15:18 +0800)]
CANN: Support operator SIN COS ARGMAX (#12709)

* [CANN]support sin cos argmax

Signed-off-by: noemotiovon <redacted>
* [CANN]codestyle adjustment

Signed-off-by: noemotiovon <redacted>
* [CANN]Remove redundant code

Signed-off-by: noemotiovon <redacted>
---------

Signed-off-by: noemotiovon <redacted>
Co-authored-by: noemotiovon <redacted>
2 months agoSimplify and improve CUDA graphs through use of indirect copy pointers (#9017)
Alan Gray [Thu, 3 Apr 2025 01:31:15 +0000 (02:31 +0100)]
Simplify and improve CUDA graphs through use of indirect copy pointers (#9017)

* CUDA: Simplify and improve CUDA graphs through use of indirect copy pointers

Previously there was complexity in the CUDA graphs implementation due
frequently changing parameters to copy kernels associated with K and V
cache pointers. This patch simplifies by using indirection to avoid
such parameters frequently changing, avoiding the need for frequent
graph updates.

Fixes #12152

* Addressed comments

* fix HIP builds

* properly sync to stream

* removed ggml_cuda_cpy_fn_ptrs

* move stream sync before free

* guard to only use indirection with graphs

* style fixes

* check for errors

---------

Co-authored-by: slaren <redacted>
2 months agoCANN: Fix failed test cases (#12708)
hipudding [Thu, 3 Apr 2025 00:49:51 +0000 (08:49 +0800)]
CANN: Fix failed test cases (#12708)

* CANN: Fix memory waste in aclnn_tensor

* CANN: fix backend ops fail

* CANN: fix acl_tensor memory alloc.

* CANN: format

* CANN: remove trailing whitespace

2 months agoopencl: use `max_alloc_size` in backend ctx instead of querying again (#12705)
lhez [Thu, 3 Apr 2025 00:01:42 +0000 (17:01 -0700)]
opencl: use `max_alloc_size` in backend ctx instead of querying again (#12705)

2 months agovulkan: Implement split_k for coopmat2 flash attention. (#12627)
Jeff Bolz [Wed, 2 Apr 2025 19:25:08 +0000 (14:25 -0500)]
vulkan: Implement split_k for coopmat2 flash attention. (#12627)

When using group query attention, we have one workgroup per KV batch and this
can be very few workgroups (e.g. just 8 in some models). Enable split_k to
spread the work across SMs. This helps a lot when the KV cache is large.

2 months agocmake: remove caching from vulkan coopmat checks (#12719)
bandoti [Wed, 2 Apr 2025 17:56:26 +0000 (14:56 -0300)]
cmake: remove caching from vulkan coopmat checks (#12719)

2 months agovulkan: Implement grouped query attention in the coopmat2 FA shader (#12559)
Jeff Bolz [Wed, 2 Apr 2025 17:40:32 +0000 (12:40 -0500)]
vulkan: Implement grouped query attention in the coopmat2 FA shader (#12559)

When adjacent batches of Q share the same batches of K/V, batch them into
the same workgroup. For example, when:

dst(128,32,1,1) = FA(q(128,1,32,1), k(128,16640,8,1), v(128,16640,8,1))

previously we would run 32 workgroups computing 1 result each, now we will
run 8 workgroups computing 4 results each.

This doesn't directly translate to better performance (at least when you have
>=32 SMs), but in a subsequent change I'll enable split_k which will scale much
better with 4x fewer workgroups.

2 months agoVulkan: Fix mmq int dot float cache size (#12722)
0cc4m [Wed, 2 Apr 2025 17:12:30 +0000 (19:12 +0200)]
Vulkan: Fix mmq int dot float cache size (#12722)

2 months agomodel : print tensor size during load (#12711)
Georgi Gerganov [Wed, 2 Apr 2025 13:38:54 +0000 (16:38 +0300)]
model : print tensor size during load (#12711)

* model : print tensor size during load

* cont : fix units MB -> MiB

Co-authored-by: Diego Devesa <redacted>
---------

Co-authored-by: Diego Devesa <redacted>
2 months agollama : add option to override model tensor buffers (#11397) upstream/0.0.5028
Diego Devesa [Wed, 2 Apr 2025 12:52:01 +0000 (14:52 +0200)]
llama : add option to override model tensor buffers (#11397)

* llama : add option to override tensor buffers

* ggml : fix possible underflow in ggml_nbytes

2 months agollama : refactor kv cache guard (#12695)
Georgi Gerganov [Wed, 2 Apr 2025 11:32:59 +0000 (14:32 +0300)]
llama : refactor kv cache guard (#12695)

* llama : refactor kv cache guard

ggml-ci

* cont : fix comment [no ci]

* llama : fix kv_cache restore logic

ggml-ci

* context : simplify kv cache updates

ggml-ci

* cont : better name [no ci]

* llama : fix llama_decode return code when could not find KV slot

ggml-ci

* context : change log err -> warn [no ci]

* kv-cache : add comment + warning

2 months agovocab : BailingMoE : change possessive quantifiers to greedy (#12677)
Sigbjørn Skjæret [Wed, 2 Apr 2025 09:21:48 +0000 (11:21 +0200)]
vocab : BailingMoE : change possessive quantifiers to greedy (#12677)

2 months agocommon : remove json.hpp from common.cpp (#12697)
Xuan-Son Nguyen [Wed, 2 Apr 2025 07:58:34 +0000 (09:58 +0200)]
common : remove json.hpp from common.cpp (#12697)

* common : remove json.hpp from common.cpp

* fix comment

2 months ago[CANN] get_rows and dup optimization (#12671)
Chenguang Li [Wed, 2 Apr 2025 07:22:13 +0000 (15:22 +0800)]
[CANN] get_rows and dup optimization (#12671)

* [CANN]get_rows and dup optimization.

Co-authored-by: hipudding <redacted>
Signed-off-by: noemotiovon <redacted>
* [CANN]GET_ROWS and CPY/DUP optimization

Co-authored-by: hipudding <redacted>
Signed-off-by: noemotiovon <redacted>
* [CANN]code style adjustment

Signed-off-by: noemotiovon <redacted>
* [CANN]code style adjustment

Signed-off-by: noemotiovon <redacted>
* [CANN]code style adjustment

Signed-off-by: noemotiovon <redacted>
* [CANN]code style adjustment

Signed-off-by: noemotiovon <redacted>
---------

Signed-off-by: noemotiovon <redacted>
Co-authored-by: noemotiovon <redacted>
Co-authored-by: hipudding <redacted>
2 months agocommon : refactor downloading system, handle mmproj with -hf option (#12694)
Xuan-Son Nguyen [Tue, 1 Apr 2025 21:44:05 +0000 (23:44 +0200)]
common : refactor downloading system, handle mmproj with -hf option (#12694)

* (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

2 months agoopencl : fix memory allocation size (#12649)
Junil Kim [Tue, 1 Apr 2025 16:54:34 +0000 (01:54 +0900)]
opencl : fix memory allocation size (#12649)

issue:
https://github.com/CodeLinaro/llama.cpp/pull/17#issuecomment-2760611283

This patch fixes the memory allocation size
not exceeding the maximum size of the OpenCL device.

2 months agollama : use LLM_KV_GENERAL_FILE_TYPE instead of gguf_find_key (#12672)
jklincn [Tue, 1 Apr 2025 12:54:28 +0000 (20:54 +0800)]
llama : use LLM_KV_GENERAL_FILE_TYPE instead of gguf_find_key (#12672)

2 months agoconvert : BailingMoE : fix qkv split when head_dim is 0 (#12687)
Sigbjørn Skjæret [Tue, 1 Apr 2025 12:37:13 +0000 (14:37 +0200)]
convert : BailingMoE : fix qkv split when head_dim is 0 (#12687)

NOTE: Ling-lite-base is broken, see https://huggingface.co/inclusionAI/Ling-lite-base/discussions/2

2 months agometal : use F32 prec in FA kernels (#12688)
Georgi Gerganov [Tue, 1 Apr 2025 11:57:19 +0000 (14:57 +0300)]
metal : use F32 prec in FA kernels (#12688)

* metal : use F32 prec in FA kernels

ggml-ci

* cont : fix FA vec kernel

ggml-ci

2 months agoFix clang warning in gguf_check_reserved_keys (#12686)
R0CKSTAR [Tue, 1 Apr 2025 11:12:53 +0000 (19:12 +0800)]
Fix clang warning in gguf_check_reserved_keys (#12686)

* Fix clang warning in gguf_check_reserved_keys

Signed-off-by: Xiaodong Ye <redacted>
* Fix typo

Signed-off-by: Xiaodong Ye <redacted>
---------

Signed-off-by: Xiaodong Ye <redacted>
2 months agovulkan: fix build when glslc doesn't support coopmat (#12683)
Wagner Bruna [Tue, 1 Apr 2025 09:38:07 +0000 (06:38 -0300)]
vulkan: fix build when glslc doesn't support coopmat (#12683)

2 months agoSYCL: Rename oneMKL to oneMath (#12192)
Romain Biessy [Tue, 1 Apr 2025 08:24:29 +0000 (10:24 +0200)]
SYCL: Rename oneMKL to oneMath (#12192)

* Rename oneMKL Interface to oneMath

* Use oneMath for Intel vendor

* Rename occurences to mkl

* clang-format

* Silence verbose warnings

* Set oneMath HIP_TARGETS

* Fix silence warnings

* Remove step to build oneMath from build instructions

* Use fixed oneMath version

* Remove INTEL_CPU

* Fold CMake oneDNN conditions

* Use Intel oneMKL for Intel devices

* Improve CMake message

* Link against MKL::MKL_SYCL::BLAS only

* Move oneMath documentation to Nvidia and AMD sections

2 months agoSYCL: switch to SYCL namespace (#12674)
Akarshan Biswas [Tue, 1 Apr 2025 08:11:39 +0000 (13:41 +0530)]
SYCL: switch to SYCL namespace (#12674)

2 months agoconvert : BailingMoE : avoid setting rope_dim to 0 (#12678)
Sigbjørn Skjæret [Mon, 31 Mar 2025 21:09:48 +0000 (23:09 +0200)]
convert : BailingMoE : avoid setting rope_dim to 0 (#12678)

2 months agovocab : add special infill tokens for CodeLlama (#11850)
Daniel Bevenius [Mon, 31 Mar 2025 16:40:56 +0000 (18:40 +0200)]
vocab : add special infill tokens for CodeLlama (#11850)

* vocab : add special infill tokens for CodeLlama

The commit adds the following special tokens for CodeLlama infill:
- `▁<PRE>`
- `▁<SUF>`
- `▁<MID>`

The motivation for this is that currently the infill example uses
CodeLlama as a suggested model. But when using this model the following
error is generated:
```console
/llama.cpp-debug/examples/infill/infill.cpp:165: GGML_ASSERT(llama_vocab_fim_pre(vocab) >= 0) failed

Could not attach to process.  If your uid matches the uid of the target
process, check the setting of /proc/sys/kernel/yama/ptrace_scope, or try
again as the root user.  For more details, see /etc/sysctl.d/10-ptrace.conf
ptrace: Operation not permitted.
No stack.
The program is not being run.
305251 Aborted                 (core dumped)
./build/bin/llama-infill -t 10 -ngl 0 -m models/codellama-13b.Q5_K_S.gguf \
  -c 4096 --temp 0.7 --repeat_penalty 1.1 -n 20 \
  --in-prefix "def helloworld():\n    print(\"hell" \
  --in-suffix "\n   print(\"goodbye world\")\n    "
```

* squash! vocab : add special infill tokens for CodeLlama

Add _<EOT> as well.

2 months agoggml : faster ssm scan (#10558)
a3sh [Mon, 31 Mar 2025 16:05:13 +0000 (00:05 +0800)]
ggml : faster ssm scan (#10558)

* faster ssm_scan

* delete unused commnet

* clang format

* add space

* modify unnecessary calculations

* faster ssm conv implementatioin

* modify file name with dash

2 months agoconvert : Qwerky : use lora_rank_tokenshift and lora_rank_decay if present (#12667)
Sigbjørn Skjæret [Mon, 31 Mar 2025 14:36:25 +0000 (16:36 +0200)]
convert : Qwerky : use lora_rank_tokenshift and lora_rank_decay if present (#12667)

2 months agoVulkan: Add DP4A MMQ and Q8_1 quantization shader (#12135)
0cc4m [Mon, 31 Mar 2025 12:37:01 +0000 (14:37 +0200)]
Vulkan: Add DP4A MMQ and Q8_1 quantization shader (#12135)

* Vulkan: Add DP4A MMQ and Q8_1 quantization shader

* Add q4_0 x q8_1 matrix matrix multiplication support

* Vulkan: Add int8 coopmat MMQ support

* Vulkan: Add q4_1, q5_0 and q5_1 quants, improve integer dot code

* Add GL_EXT_integer_dot_product check

* Remove ggml changes, fix mmq pipeline picker

* Remove ggml changes, restore Intel coopmat behaviour

* Fix glsl compile attempt when integer vec dot is not supported

* Remove redundant code, use non-saturating integer dot, enable all matmul sizes for mmq

* Remove redundant comment

* Fix integer dot check

* Fix compile issue with unsupported int dot glslc

* Update Windows build Vulkan SDK version

2 months agocmake : fix whitespace (#0)
Georgi Gerganov [Mon, 31 Mar 2025 12:05:30 +0000 (15:05 +0300)]
cmake : fix whitespace (#0)

2 months agosync : ggml
Georgi Gerganov [Mon, 31 Mar 2025 11:59:21 +0000 (14:59 +0300)]
sync : ggml

ggml-ci

2 months agocmake: improve Vulkan cooperative matrix support checks (whisper/2966)
Sandro Hanea [Mon, 31 Mar 2025 10:44:36 +0000 (12:44 +0200)]
cmake: improve Vulkan cooperative matrix support checks (whisper/2966)

Co-authored-by: Sandro Hanea <redacted>
2 months agollava : proper description fix (#12668)
Sigbjørn Skjæret [Mon, 31 Mar 2025 09:28:30 +0000 (11:28 +0200)]
llava : proper description fix (#12668)

2 months agoSYCL: Remove misleading ggml_sycl_op_flatten function (#12387)
Akarshan Biswas [Mon, 31 Mar 2025 09:25:24 +0000 (14:55 +0530)]
SYCL: Remove misleading ggml_sycl_op_flatten function (#12387)

* SYCL: Remove misleading ggml_sycl_op_flatten function

* remove trailing whitespace

* Fix L2 norm from rebase

* remove try catch block from element_wise.cpp

* remove comment from common.hp

* ggml-sycl.cpp: Add try catch sycl::exception block in compute_forward

* norm.cpp: remove try catch exception block