]>
git.djapps.eu Git - pkg/ggml/sources/llama.cpp/log
Aleksander Grygier [Wed, 12 Nov 2025 18:01:48 +0000 (19:01 +0100)]
Update packages + upgrade Storybook to v10 (#17201)
* chore: Update packages + upgrade Storybook to v10
* fix: Increase timeout for UI tests
Xuan-Son Nguyen [Wed, 12 Nov 2025 17:50:52 +0000 (18:50 +0100)]
server: (refactor) implement generator-based API for task results (#17174)
* server: (refactor) implement generator-based API for task results
* improve
* moving some code
* fix "Response ended prematurely"
* add sink.done before return false
* rm redundant check
* rm unused var
* rename generator --> reader
Xuan-Son Nguyen [Wed, 12 Nov 2025 13:56:02 +0000 (14:56 +0100)]
ci: add check vendor job (#17179)
* ci: add check vendor job
* use dev version of miniaudio
* move to dedicated workflow, only run on related files changed
Xuan-Son Nguyen [Wed, 12 Nov 2025 13:17:24 +0000 (14:17 +0100)]
server: move res_error/res_ok to static function (#17167)
Alberto Cabrera Pérez [Wed, 12 Nov 2025 12:52:19 +0000 (12:52 +0000)]
ggml-cpu: handle 3d tensors in repack mat_mul (#17030)
* ggml-cpu: handle 3d tensors in repack mul_mat
* Removed unnecessary branch, removed need for <algorithm>
* Fixed dst_ptr pointer in chunk + clang_format
* GGML_ASSERT to check wdata within bounds
* Accidental ggml.h inclusion
* Improved GGML_ASSERT on wdata boundaries
Adrien Gallouët [Wed, 12 Nov 2025 12:48:30 +0000 (13:48 +0100)]
cmake : cleanup (#17199)
Adrien Gallouët [Wed, 12 Nov 2025 11:32:50 +0000 (12:32 +0100)]
cmake : move OpenSSL linking to vendor/cpp-httplib (#17177)
* cmake : move OpenSSL linking to vendor/cpp-httplib
Signed-off-by: Adrien Gallouët <redacted>
* bring back httplib 0.27.0
* add -DLLAMA_HTTPLIB
* update cmake config for visionos
---------
Signed-off-by: Adrien Gallouët <redacted>
Co-authored-by: Xuan Son Nguyen <redacted>
TecJesh [Wed, 12 Nov 2025 07:11:42 +0000 (15:11 +0800)]
CANN: Add L2_NORM op support (#16856)
* update L2_NORM op support
* update L2_NORM op support
* remove extra whitespace
Neo Zhang Jianyu [Wed, 12 Nov 2025 06:44:29 +0000 (14:44 +0800)]
[SYCL]fix ci crash about SSM_CONV (#17169)
* fix ci crash
* Update ggml-sycl.cpp
* Update ggml/src/ggml-sycl/ggml-sycl.cpp
Co-authored-by: Sigbjørn Skjæret <redacted>
---------
Co-authored-by: Zhang Jianyu <redacted>
Co-authored-by: Sigbjørn Skjæret <redacted>
Raul Torres [Wed, 12 Nov 2025 06:37:52 +0000 (06:37 +0000)]
CANN: GGML_CANN_ACL_GRAPH works only USE_ACL_GRAPH enabled (#16861)
The documentation should state that `GGML_CANN_ACL_GRAPH` is only effective if `USE_ACL_GRAPH` was enabled at compilation time.
Max Krasnyansky [Tue, 11 Nov 2025 23:25:04 +0000 (15:25 -0800)]
hexagon: various Op fixes (#17135)
* hexagon: explicitly check for ops with zero nrows
llm_graph_context::build_inp_out_ids() can generate tensors with zero nrows.
Somehow other backends seems to handle this without obvious explicit checks.
In the hexagon case we need to check explicitly and skip them.
* hexagon: introduce fastdiv, fix test-backend-ops for ADD/SUB/MUL
Co-authored-by: chraac <redacted>
* hexagon: use fastdiv in ADD_ID
* hexagon: use ggml_op_is_empty and ggml_is_empty to check for NOPs
---------
Co-authored-by: chraac <redacted>
Eve [Tue, 11 Nov 2025 18:53:30 +0000 (18:53 +0000)]
disable rms norm mul rope for chips with no fp16 rte (#17134)
sudhiarm [Tue, 11 Nov 2025 15:58:05 +0000 (15:58 +0000)]
ci: add Arm-hosted Graviton4 runner (#17021)
* ci: add Arm-hosted Graviton4 runner
* ci: add missing dependencies for graviton4 build
* ci: enable LFS checkout on graviton4
* ci: move git-lfs install to dependencies in Graviton4 workflow
Xuan-Son Nguyen [Tue, 11 Nov 2025 12:32:58 +0000 (13:32 +0100)]
vendor: split httplib to cpp/h files (#17150)
* vendor: split httplib to cpp/h files
* move defines
* include httplib if curl is not used
* add TODO
* fix build ios
* fix build visionos instead
ixgbe [Tue, 11 Nov 2025 11:41:51 +0000 (19:41 +0800)]
ggml-cpu : add RISC-V RVV (Zvfh) optimization for FP16 to FP32 conversion (#17161)
Signed-off-by: Wang Yang <redacted>
duduta [Tue, 11 Nov 2025 11:33:24 +0000 (13:33 +0200)]
ggml-cpu: templateify ggml_compute_forward_rope_f32 and _f16 (#16805)
* extract rotate_pairs logic from ggml_compute_forward_rope_f32
* templateify ggml_compute_forward_rope_f32 and _f16
* abort when rope type not supported, remove GLM from test-rope
* add imrope branch to switch
* add rope tests for perf
* Update ggml/src/ggml-cpu/ops.cpp
Co-authored-by: Georgi Gerganov <redacted>
* Update ggml/src/ggml-cpu/ops.cpp
Co-authored-by: Georgi Gerganov <redacted>
---------
Co-authored-by: Georgi Gerganov <redacted>
Charles Xu [Tue, 11 Nov 2025 11:20:31 +0000 (12:20 +0100)]
kleidiai: add optimized per-channel kernels for Q8_0 (#16993)
Mike Abbott [Tue, 11 Nov 2025 11:19:50 +0000 (04:19 -0700)]
cmake : add version to all shared object files (#17091)
When compiling llama.cpp in Yocto, it fails QA checks because the generated so files aren't versioned. This applies a version to all generated so files, allowing the package to build without errors.
Nicolas B. Pierron [Tue, 11 Nov 2025 10:53:59 +0000 (11:53 +0100)]
Install rpc-server when GGML_RPC is ON. (#17149)
levkropp [Tue, 11 Nov 2025 08:38:30 +0000 (03:38 -0500)]
convert : register UMT5Model architecture for T5 conversion (#17160)
Register UMT5Model as a supported architecture variant for T5 model conversion.
This allows the conversion to work for models downloaded with AutoModel.
lhez [Mon, 10 Nov 2025 23:00:13 +0000 (15:00 -0800)]
opencl: add fastdiv and use it in set_rows, ported from cuda (#17090)
* opencl: add fastdiv for mm q8_0
* opencl: use uint4 for fastdiv vals
* opencl: use fastdiv for set_rows
* opencl: do not use fastdiv for q8_0 mm
Sigbjørn Skjæret [Mon, 10 Nov 2025 21:55:30 +0000 (22:55 +0100)]
models : move build_inp_out_ids outside loop (#17151)
* move build_inp_out_ids outside loop
* realign
Max Krasnyansky [Mon, 10 Nov 2025 20:44:49 +0000 (12:44 -0800)]
cpu: skip NOPs to avoid barriers (#17133)
* cpu: skip NOPs to avoid barriers
* cpu: use ggml_op_is_empty
Georgi Gerganov [Mon, 10 Nov 2025 19:33:35 +0000 (21:33 +0200)]
metal : cap threadgroups size of set_rows (#17146)
Adrien Gallouët [Mon, 10 Nov 2025 19:03:36 +0000 (20:03 +0100)]
ggml-cpu : inspect -march and -mcpu to found the CPU (#16333)
Signed-off-by: Adrien Gallouët <redacted>
Ruben Ortlam [Mon, 10 Nov 2025 15:59:26 +0000 (16:59 +0100)]
vulkan: check glslc executable string (#17144)
Ruben Ortlam [Mon, 10 Nov 2025 15:59:10 +0000 (16:59 +0100)]
vulkan: fix validation issue introduced by #16868 (#17145)
Gabe Goodhart [Mon, 10 Nov 2025 15:14:23 +0000 (08:14 -0700)]
memory: Hybrid context shift (#17009)
* feat(memory): Only fail partial erasure of recurrent tail
The recurrent state is always assumed to be the state as of the last update
from the final token in the sequence. When doing a partial erasure, if the
range does not include the final token, the erasure can be considered a
success since any memory used for the sequence prior to the final token
(which is no memory) has been successfully removed.
There is one potential case that this doesn't address which is the pruning
of cache to remove sensitive data from the context. This wouldn't work for
attention cache partial removal (in the middle) either since the KV state
is linearly-dependent and states in later sequence positions would still be
based on the state from the sensitive data, even if that data is no longer
cached, so I don't think this is relevant, but it is worth noting that the
semantics of this change for a partial erasure in the middle of the cache
are essentially "my context is already compressed" and not "all trace of
the removed tokens has been removed."
https://github.com/ggml-org/llama.cpp/issues/16768
Branch: HybridContextShift-16768
Signed-off-by: Gabe Goodhart <redacted>
* fix(main): Check the output of seq_rm for prefix matching
This prefix matching is explicitly attempting to remove the tokens at the
end of the sequence that don't match. This is the operation that can't be
performed on a recurrent cache due to the state being updated in place, so
if this removal fails, we need to clear the whole cache.
https://github.com/ggml-org/llama.cpp/issues/16768
Branch: HybridContextShift-16768
Signed-off-by: Gabe Goodhart <redacted>
* fix(memory): Fix condition for partial erasure failure if p0 > pos
Signed-off-by: Gabe Goodhart <redacted>
Co-authored-by: compilade <redacted>
* style: Fix extra parens
Signed-off-by: Gabe Goodhart <redacted>
Co-authored-by: Georgi Gerganov <redacted>
* fix(main.cpp): Set n_matching_session_tokens to 0 on cache clear
https://github.com/ggml-org/llama.cpp/issues/16768
Branch: HybridContextShift-16768
Signed-off-by: Gabe Goodhart <redacted>
---------
Signed-off-by: Gabe Goodhart <redacted>
Co-authored-by: compilade <redacted>
Co-authored-by: Georgi Gerganov <redacted>
Georgi Gerganov [Mon, 10 Nov 2025 13:38:42 +0000 (15:38 +0200)]
metal : enable tensor API for A19 (#17087)
fj-y-saito [Mon, 10 Nov 2025 13:12:59 +0000 (22:12 +0900)]
arm64: add i8mm route with SVE ggml_vec_dot_q4_K_q8_K and ggml_vec_dot_q6_K_… (#15277)
* add i8mm route with SVE ggml_vec_dot_q4_K_q8_K and ggml_vec_dot_q6_K_q8_K
* Surround SVE function with compiler directive
* fix compile switch
* fix coding style
* ggml : fix indent
---------
Co-authored-by: Georgi Gerganov <redacted>
Georgi Gerganov [Mon, 10 Nov 2025 10:59:29 +0000 (12:59 +0200)]
batched-bench : add "separate text gen" mode (#17103)
Xuan-Son Nguyen [Mon, 10 Nov 2025 10:41:05 +0000 (11:41 +0100)]
mtmd: fix patch_size initialized to random value in audio models (#17128)
* mtmd: fix patch_size initialized to random value in audio models
* add default hparams
Georgi Gerganov [Mon, 10 Nov 2025 10:17:19 +0000 (12:17 +0200)]
editorconfig : ignore benches/ (#17140)
[no ci]
Acly [Mon, 10 Nov 2025 09:19:39 +0000 (10:19 +0100)]
cuda/vulkan : bicubic interpolation (#17022)
* vulkan : implement upscale with bicubic interpolation
* cuda : implement upscale with bicubic interpolation
* tests : add ggml_interpolate with GGML_SCALE_MODE_BICUBIC to backend tests
* adapt OpenCL backend to not support the OP in that case so tests don't fail
* print scale mode & flags in test-backend-ops
Georgi Gerganov [Mon, 10 Nov 2025 08:44:10 +0000 (10:44 +0200)]
benches : add eval results (#17139)
[no ci]
Georgi Gerganov [Sun, 9 Nov 2025 16:31:02 +0000 (18:31 +0200)]
mtmd : fix embedding size for image input (#17123)
Ruben Ortlam [Sun, 9 Nov 2025 15:14:41 +0000 (16:14 +0100)]
vulkan: fix memory allocations (#17122)
compilade [Sun, 9 Nov 2025 14:49:40 +0000 (09:49 -0500)]
convert : parse safetensors directly (#15667)
* convert : parse safetensors directly
* gguf-py : order safetensors tensors by name
Applies to both local and remote safetensors custom parsing.
This matches the behavior of the official safetensors implementation.
* convert : rename from_safetensors_meta to from_local_tensor
For consistency with from_remote_tensor
* convert : fix no-lazy dtypes from direct safetensors
compilade [Sun, 9 Nov 2025 14:45:50 +0000 (09:45 -0500)]
convert : handle compressed-tensors quant method (#17069)
* convert : handle compressed-tensors quant method
* convert : handle int-quantized models
* convert : handle naive-quantized models
* gguf-py : __pos__ is also unary
* convert : fix flake8 lint
* convert : use F32 for dequant of pack-quantized tensors
Georgi Gerganov [Sun, 9 Nov 2025 12:27:05 +0000 (14:27 +0200)]
server : handle failures to restore host cache (#17078)
* server : handle failures to restore host cache
* server : add tests for the prompt cache
Georgi Gerganov [Sun, 9 Nov 2025 10:53:29 +0000 (12:53 +0200)]
benches : add folder with benchmarks (#16931)
* benches : add folder with benchmarks
* benches : update dgx-spark bench
Eric Curtin [Sun, 9 Nov 2025 09:25:38 +0000 (09:25 +0000)]
Switch to using Ubuntu 25.10 vulkan/mesa (#16497)
Because "Ubuntu packages to be discontinued in Vulkan SDK"
Signed-off-by: Eric Curtin <redacted>
Ruben Ortlam [Sun, 9 Nov 2025 08:54:47 +0000 (09:54 +0100)]
vulkan: iGPU memory reporting fix (#17110)
* vulkan: use all device-local heaps for memory availability reporting
Co-authored-by: Giuseppe Scrivano <redacted>
* use all available heaps for iGPU memory reporting
* Allow multiple memory types per buffer request for devices with split heaps
---------
Co-authored-by: Giuseppe Scrivano <redacted>
Ruben Ortlam [Sun, 9 Nov 2025 08:52:57 +0000 (09:52 +0100)]
vulkan: fix mmq out of bounds reads (#17108)
* vulkan: fix mmq out of bounds reads, streamline outdated matmul host code
* fix mul_mat_id quantization call
* Fix compiler warnings
Jeff Bolz [Sun, 9 Nov 2025 08:48:42 +0000 (02:48 -0600)]
vulkan: fuse mul_mat_id + mul (#17095)
* vulkan: fuse mul_mat_id + mul
This comes up in qwen3 moe.
* split mul_mat_id fusion tests into a separate class
Georgi Gerganov [Sun, 9 Nov 2025 06:28:51 +0000 (08:28 +0200)]
metal : retain src and dst buffers during async ops (#17101)
Xuan-Son Nguyen [Sat, 8 Nov 2025 20:54:14 +0000 (21:54 +0100)]
arg: add --cache-list argument to list cached models (#17073)
* arg: add --cache-list argument to list cached models
* new manifest naming format
* improve naming
* Update common/arg.cpp
Co-authored-by: Georgi Gerganov <redacted>
---------
Co-authored-by: Georgi Gerganov <redacted>
chansikpark [Sat, 8 Nov 2025 19:52:35 +0000 (14:52 -0500)]
webui: fix keyboard shortcuts for new chat & edit chat title (#17007)
Jeff Bolz [Sat, 8 Nov 2025 19:24:29 +0000 (13:24 -0600)]
vulkan: Use spec constants for conv2d s/d/p and kernel W/H (#16978)
* vulkan: Use spec constants for conv2d s/d/p and kernel W/H
Also add some additional unroll hints, which seems to help.
* lock around map lookup
Aidan [Sat, 8 Nov 2025 13:12:11 +0000 (13:12 +0000)]
server: fix correct time_ms calculation in prompt_progress (#17093)
* fix: correct time_ms calculation in send_partial_response
The time_ms field was incorrectly calculated. The division was happening
before the subtraction leading to incorrect values.
Before: (ggml_time_us() - slot.t_start_process_prompt / 1000) After:
(ggml_time_us() - slot.t_start_process_prompt) / 1000
* docs : document time_ms field in prompt_progress
Aman Gupta [Sat, 8 Nov 2025 13:05:19 +0000 (21:05 +0800)]
Revert "CUDA: add expert reduce kernel (#16857)" (#17100)
Aman Gupta [Sat, 8 Nov 2025 08:58:05 +0000 (16:58 +0800)]
CUDA: skip fusion for repeating adds in bias (#17080)
SavicStefan [Sat, 8 Nov 2025 08:28:22 +0000 (09:28 +0100)]
vulkan: Increase BK to 32; use BK/4 for non-CM mul_mm.comp (#16636)
Signed-off-by: Stefan Savic <redacted>
Co-authored-by: Stefan Savic <redacted>
Aleksei Nikiforov [Sat, 8 Nov 2025 08:00:20 +0000 (09:00 +0100)]
ggml: disable vxe for cross-compilation by default (#16966)
Otherwise compilation will fail due to enabling -mvx -mzvector
and not setting corresponding -march options.
Jeff Bolz [Sat, 8 Nov 2025 07:52:15 +0000 (01:52 -0600)]
vulkan: fuse rms_norm + mul + rope (+ view + set_rows) (#16977)
This change combines the rms_norm+mul and rope+view+set_rows fusions to
allow fusing the whole sequence together. This comes up in Qwen3, Bailing,
and some other models.
Jeff Bolz [Sat, 8 Nov 2025 07:39:45 +0000 (01:39 -0600)]
vulkan: Fix test-thread-safety crashes (#17024)
The std::map pipeline_flash_attn_f32_f16 could be searched and inserted at the
same time, which needs to hold the lock. To be safe, hold the lock for all of
ggml_vk_load_shaders.
Johannes Gäßler [Sat, 8 Nov 2025 07:26:18 +0000 (08:26 +0100)]
CUDA: fix MMQ stream-k fixup ne1 indices (#17089)
Reese Levine [Sat, 8 Nov 2025 03:27:20 +0000 (19:27 -0800)]
ggml webgpu: faster matrix multiplication/matrix-vector multiplication (#17031)
* Faster tensors (#8)
Add fast matrix and matrix/vector multiplication.
* Use map for shader replacements instead of pair of strings
bssrdf [Fri, 7 Nov 2025 22:41:58 +0000 (17:41 -0500)]
CUDA: properly handle nb00=nb02 case for cpy (#17081)
Acly [Fri, 7 Nov 2025 20:08:50 +0000 (21:08 +0100)]
vulkan : refactor buffer handling in vk_op_f32 (#16840)
* vulkan : refactor/simplify buffer handling in vk_op_* functions
* Combine UMA handling into ggml_vk_tensor_subbuffer
Johannes Gäßler [Fri, 7 Nov 2025 19:53:14 +0000 (20:53 +0100)]
CUDA: fix should_use_mmvf for ne11 == 1 (#17085)
* CUDA: fix should_use_mmvf for ne11 == 1
* Apply suggestion from @am17an
Co-authored-by: Aman Gupta <redacted>
---------
Co-authored-by: Aman Gupta <redacted>
Georgi Gerganov [Fri, 7 Nov 2025 19:23:11 +0000 (21:23 +0200)]
bench : cache the llama_context state at computed depth (#16944)
* bench : cache llama_context state at depth
* cont : handle failures to restore the old state
* cont : print information when the state is being reused
Sigbjørn Skjæret [Fri, 7 Nov 2025 18:27:58 +0000 (19:27 +0100)]
hparams : add n_embd_inp() to support extended embed (#16928)
* add n_embd_full to support extended embed
* don't change output
* rename to n_embd_inp
* restore n_embd where applicable
Georgi Gerganov [Fri, 7 Nov 2025 18:03:25 +0000 (20:03 +0200)]
kv-cache : pad the cache size to 256 for performance (#17046)
* kv-cache : pad the size of the small SWA cache for performance
* context : pad the total context to 256
* cont : future-proof the swa pad
* server : adjust test params to new logic
Adrien Gallouët [Fri, 7 Nov 2025 16:34:05 +0000 (17:34 +0100)]
Revert "ggml-cpu: detect correct cpu flags for arm64 (#16229) (#16239)" (#17084)
This reverts commit
7c23f3f0d4b9f5d6ea140756eb694b562d5acebb .
iron [Fri, 7 Nov 2025 16:18:14 +0000 (00:18 +0800)]
ggml-cpu: detect correct cpu flags for arm64 (#16229) (#16239)
When using GCC 9 and GCC 12 on the arm64 platform of ubuntu 2004,
the command "gcc -mcpu=native -E -v -" fails to detect the correct CPU flags,
which results in compilation failures for certain extended instructions,
but the correct CPU flags can be obtained by using gcc -march.
Signed-off-by: lizhenneng <redacted>
Co-authored-by: lizhenneng <redacted>
Georgi Gerganov [Fri, 7 Nov 2025 10:24:47 +0000 (12:24 +0200)]
server : print the samplers chain for each request (#17070)
Xuan-Son Nguyen [Fri, 7 Nov 2025 10:23:34 +0000 (11:23 +0100)]
common: move download functions to download.(cpp|h) (#17059)
* common: move download functions to download.(cpp|h)
* rm unused includes
* minor cleanup
---------
Co-authored-by: Georgi Gerganov <redacted>
xctan [Thu, 6 Nov 2025 16:12:45 +0000 (00:12 +0800)]
ggml-cpu : optimize RVV q2_k and q3_k kernels (#16887)
Johannes Gäßler [Thu, 6 Nov 2025 13:05:47 +0000 (14:05 +0100)]
CUDA: fix crash on uneven context without FA (#16988)
Georgi Gerganov [Thu, 6 Nov 2025 12:45:10 +0000 (14:45 +0200)]
metal : initial Metal4 tensor API support (#16634)
* metal : rework mat-mat multiplication
* metal : initial Metal4 support
* cont
* metal : detect tensor support
* cont : better ifdefs
* metal : support tensors in mul_mm_id
* metal : add env for disabling tensor API
* tests : restore
* metal : remove unused constants
* metal : fix check for bfloat tensor support
* cont : handle API incompatibilities
* cont : handle even more incompatibilities
* metal : use tensor API only on M5 and later
Georgi Gerganov [Thu, 6 Nov 2025 10:09:29 +0000 (12:09 +0200)]
server : disable checkpoints with mtmd (#17045)
Xuan-Son Nguyen [Thu, 6 Nov 2025 10:02:54 +0000 (11:02 +0100)]
clip: implement minicpm-v sinusoidal embd using GGML (#17036)
* clip: implement minicpm-v sinusoidal embd using GGML
* fix repeat op
YehuditE [Thu, 6 Nov 2025 10:02:33 +0000 (12:02 +0200)]
sycl: add CONCAT operator support (#16047)
* sycl: add CONCAT operator support
* cleanup: remove stray lines added by mistake
* fix: code format issues in concat.cpp and tests/test-backend-ops.cpp
* chore: fix editorconfig violations
* cleanup: drop unnecessary i16 type support
* docs: update sycl-csv and regenerate ops.md
* update docs/ops.md
* fix: adapt to upstream master changes after rebase
* fix: remove empty files
* fix: drop whitespace
---------
Co-authored-by: Sigbjørn Skjæret <redacted>
Johannes Gäßler [Thu, 6 Nov 2025 07:14:35 +0000 (08:14 +0100)]
docs: explain CUDA 11 compilation [no ci] (#16824)
l3utterfly [Thu, 6 Nov 2025 05:46:38 +0000 (13:46 +0800)]
ggml-hexagon: graceful fallback for older socs where rpcmem_alloc2 and FASTRPC_GET_URI is unsupported (#16987)
* support older socs where FASTRPC_GET_URI is unsupported
* added graceful fallback when FASTRPC_GET_URI call fails
* use weak symbols instead of loading libcdsprpc.so dynamically
* Add weak pragma for rpcmem_alloc2
* Remove weak declaration for rpcmem_alloc2 in ggml-hexagon.cpp
Removed weak declaration for rpcmem_alloc2.
* Enforce ndev to 1 for archs below v75
Force ndev to 1 for SoCs architectures lower than v75.
bssrdf [Wed, 5 Nov 2025 20:55:04 +0000 (15:55 -0500)]
improve CUDA cpy memory bandwidth when copying transposed tensor (#16841)
* WIP
* added a cpy kernel specific to transposed tensor which uses smem to avoid uncoalesced access; test cases also added shwoing improved memory bandwidth
* added BF16 support
* more strict check to make sure src0 is a transpose
* reformulated to handle more complicated transpose cases
* bring back 2D transpose for higher performance
* allow build on windows
* tranpose copy more shapes
* minor tweak
* final clean up
* restore some test cases
* keep only the kernel for true tranposed case; updated with review suggestions
* make CI happy
* remove headers not needed
* reduced bank conflicts for fp16 and bf16
* add missing const*
* now bank conflicts free
* use padding instead of swizzling
---------
Co-authored-by: bssrdf <redacted>
Jeff Bolz [Wed, 5 Nov 2025 18:51:03 +0000 (12:51 -0600)]
vulkan: Fix GGML_VULKAN_CHECK_RESULTS to better handle fusion (#16919)
Gabe Goodhart [Wed, 5 Nov 2025 17:58:16 +0000 (10:58 -0700)]
examples(gguf): GGUF example outputs (#17025)
* feat(llama-gguf): Print out the tensor type in llama-gguf r
Branch: Mamba2Perf
Signed-off-by: Gabe Goodhart <redacted>
* feat(off-topic): print the number of elements in tensors with llama-gguf
Branch: Mamba2SSD
Signed-off-by: Gabe Goodhart <redacted>
* style: valign
Branch: GGUFToolOutputs
Signed-off-by: Gabe Goodhart <redacted>
* Update examples/gguf/gguf.cpp
---------
Signed-off-by: Gabe Goodhart <redacted>
Co-authored-by: Georgi Gerganov <redacted>
Xuan-Son Nguyen [Wed, 5 Nov 2025 13:26:49 +0000 (14:26 +0100)]
mtmd: allow QwenVL to process larger image by default (#17020)
Georgi Gerganov [Wed, 5 Nov 2025 12:32:55 +0000 (14:32 +0200)]
server : do not default to multiple slots with speculative decoding (#17017)
* server : do not default to multiple slots with speculative decoding
* cont : fix
Xuan-Son Nguyen [Wed, 5 Nov 2025 10:26:37 +0000 (11:26 +0100)]
mtmd: improve struct initialization (#16981)
손희준 [Wed, 5 Nov 2025 10:20:28 +0000 (19:20 +0900)]
docs: Clarify the endpoint that webui uses (#17001)
Li Pengzhan [Wed, 5 Nov 2025 09:28:58 +0000 (17:28 +0800)]
model : add openPangu-Embedded (#16941)
* Model: add openPangu-Embedded
* fixed according to reviewer's comments
* fixed the chat template check condition
* Apply suggestions from code review
change the chat-template check condition and some formatting issue
Co-authored-by: Sigbjørn Skjæret <redacted>
* whitespace cleanup
---------
Co-authored-by: Sigbjørn Skjæret <redacted>
Reese Levine [Wed, 5 Nov 2025 09:27:42 +0000 (01:27 -0800)]
ggml webgpu: minor set rows optimization (#16810)
* Add buffer label and enable dawn-specific toggles to turn off some checks
* 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 <redacted>
Co-authored-by: Neha Abbas <redacted>
Co-authored-by: Reese Levine <redacted>
* Comment on dawn toggles
* Remove some comments
* Implement overlap binary operators
* Revert "Implement overlap binary operators"
This reverts commit
ed710b36f51ab3f53fa13db15c1685dc8678a32a .
* Disable support for non-contiguous binary_op tensors and leave note for future support
---------
Co-authored-by: neha-ha <redacted>
Co-authored-by: Neha Abbas <redacted>
Co-authored-by: Neha Abbas <redacted>
Georgi Gerganov [Tue, 4 Nov 2025 18:44:18 +0000 (20:44 +0200)]
sync : ggml
Georgi Gerganov [Tue, 4 Nov 2025 18:40:52 +0000 (20:40 +0200)]
ggml : fix conv2d_dw SVE path (ggml/1380)
* Fix test-conv2d-dw failure on ARM SVE by using runtime vector length
The ggml_compute_forward_conv_2d_dw_cwhn function was using a hardcoded GGML_F32_EPR (8) for SIMD vectorization, but on ARM SVE the actual vector length varies by hardware. This caused incorrect computation when processing CWHN layout tensors on ARM machines.
Fix by using svcntw() to get the runtime SVE vector length instead of the compile-time constant.
Co-authored-by: ggerganov <redacted>
* ci : reduce sam score threshold
* ci : update bbox checks for sam test
---------
Co-authored-by: copilot-swe-agent[bot] <redacted>
Co-authored-by: ggerganov <redacted>
mnehete32 [Wed, 5 Nov 2025 03:01:15 +0000 (08:31 +0530)]
CUDA: update ops.md (#17005)
lhez [Wed, 5 Nov 2025 00:02:36 +0000 (16:02 -0800)]
opencl: update doc (#17011)
* opencl: update docs
* opencl: update docs
* opencl: fix link
* opencl: update doc
nullname [Tue, 4 Nov 2025 20:25:39 +0000 (04:25 +0800)]
refactor: replace sprintf with snprintf for safer string handling in dump functions (#16913)
Jeff Bolz [Tue, 4 Nov 2025 19:28:17 +0000 (13:28 -0600)]
vulkan: remove the need for the dryrun (#16826)
* vulkan: remove the need for the dryrun
Allocate pipelines and descriptor sets when requested.
Reallocate the prealloc buffers when needed, and flush any pending work
before reallocating.
For rms_partials and total_mul_mat_bytes, use the sizes computed the last time
the graph was executed.
* remove dryrun parameters
Georgi Gerganov [Tue, 4 Nov 2025 17:21:36 +0000 (19:21 +0200)]
server : do context shift only while generating (#17000)
Georgi Gerganov [Tue, 4 Nov 2025 15:21:31 +0000 (17:21 +0200)]
readme : update hot topics (#17002)
Acly [Tue, 4 Nov 2025 12:12:20 +0000 (13:12 +0100)]
ggml-cpu : bicubic interpolation (#16891)
Sigbjørn Skjæret [Tue, 4 Nov 2025 11:29:39 +0000 (12:29 +0100)]
ci : apply model label to models (#16994)
Sigbjørn Skjæret [Tue, 4 Nov 2025 11:29:15 +0000 (12:29 +0100)]
chore : fix models indent after refactor (#16992)
Noah [Tue, 4 Nov 2025 05:04:59 +0000 (05:04 +0000)]
Fix garbled output with REPACK at high thread counts (#16956)
* Fix garbled output with REPACK at high thread counts
Fixed a race condition in the REPACK matrix multiplication code that caused garbled output when using 26+ threads (model-dependent threshold). The issue occurred because with high thread counts, the code forced chunk count to equal thread count, creating many small chunks. After aligning these chunks to NB_COLS boundaries, adjacent chunks could overlap, causing data corruption and race conditions. The fix enforces minimum chunk sizes based on NB_COLS and caps maximum chunk count to prevent creating too many tiny chunks, ensuring proper alignment without overlaps.
* Update ggml/src/ggml-cpu/repack.cpp
Co-authored-by: Georgi Gerganov <redacted>
* Update ggml/src/ggml-cpu/repack.cpp
Co-authored-by: Georgi Gerganov <redacted>
---------
Co-authored-by: Georgi Gerganov <redacted>
Aman Gupta [Tue, 4 Nov 2025 02:53:48 +0000 (10:53 +0800)]
CUDA: avoid mul + bias fusion when doing fusion (#16935)
lhez [Mon, 3 Nov 2025 19:47:57 +0000 (11:47 -0800)]
opencl: support imrope (#16914)
* opencl: support imrope
* opencl: fix whitespace
Aleksander Grygier [Mon, 3 Nov 2025 17:53:26 +0000 (18:53 +0100)]
fix: Viewing multiple PDF attachments (#16974)