]>
git.djapps.eu Git - pkg/ggml/sources/llama.cpp/log
Chenguang Li [Tue, 26 Nov 2024 09:31:05 +0000 (17:31 +0800)]
CANN: RoPE and CANCAT operator optimization (#10488)
Co-authored-by: noemotiovon <redacted>
Junil Kim [Tue, 26 Nov 2024 01:47:20 +0000 (10:47 +0900)]
vulkan: Fix a vulkan-shaders-gen arugment parsing error (#10484)
The vulkan-shaders-gen was not parsing the --no-clean argument correctly.
Because the previous code was parsing the arguments which have a value only
and the --no-clean argument does not have a value, it was not being parsed
correctly. This commit can now correctly parse arguments that don't have values.
Eric Curtin [Mon, 25 Nov 2024 21:56:24 +0000 (16:56 -0500)]
Introduce llama-run (#10291)
It's like simple-chat but it uses smart pointers to avoid manual
memory cleanups. Less memory leaks in the code now. Avoid printing
multiple dots. Split code into smaller functions. Uses no exception
handling.
Signed-off-by: Eric Curtin <redacted>
Diego Devesa [Mon, 25 Nov 2024 21:05:39 +0000 (22:05 +0100)]
ci : build docker images only once daily (#10503)
Georgi Gerganov [Mon, 25 Nov 2024 20:28:27 +0000 (22:28 +0200)]
server : add more information about error (#10455)
Georgi Gerganov [Mon, 25 Nov 2024 19:50:07 +0000 (21:50 +0200)]
server : enable cache_prompt by default (#10501)
ggml-ci
Georgi Gerganov [Mon, 25 Nov 2024 19:49:31 +0000 (21:49 +0200)]
metal : enable mat-vec kernels for bs <= 4 (#10491)
Shane A [Mon, 25 Nov 2024 18:36:09 +0000 (10:36 -0800)]
Rename Olmo1124 to Olmo2 (#10500)
Diego Devesa [Mon, 25 Nov 2024 18:30:06 +0000 (19:30 +0100)]
llama : accept a list of devices to use to offload a model (#10497)
* llama : accept a list of devices to use to offload a model
* accept `--dev none` to completely disable offloading
* fix dev list with dl backends
* rename env parameter to LLAMA_ARG_DEVICE for consistency
Johannes Gäßler [Mon, 25 Nov 2024 18:18:37 +0000 (19:18 +0100)]
Github: update issue templates [no ci] (#10489)
brucepro [Mon, 25 Nov 2024 16:11:55 +0000 (08:11 -0800)]
Add download chat feature to server chat (#10481)
* Add download chat feature to server chat
Add a download feature next to the delete chat feature in the server vue chat interface.
* code style
---------
Co-authored-by: Xuan Son Nguyen <redacted>
Georgi Gerganov [Mon, 25 Nov 2024 14:31:38 +0000 (16:31 +0200)]
server : add speculative decoding support (#10455)
* server : add speculative decoding support
ggml-ci
* server : add helper function slot.can_speculate()
ggml-ci
Diego Devesa [Mon, 25 Nov 2024 14:13:39 +0000 (15:13 +0100)]
ggml : add support for dynamic loading of backends (#10469)
* ggml : add support for dynamic loading of backends
---------
Co-authored-by: Georgi Gerganov <redacted>
Georgi Gerganov [Mon, 25 Nov 2024 13:17:32 +0000 (15:17 +0200)]
tests : fix compile warning
Georgi Gerganov [Mon, 25 Nov 2024 13:08:04 +0000 (15:08 +0200)]
metal : minor code formatting
Neo Zhang Jianyu [Mon, 25 Nov 2024 09:31:10 +0000 (17:31 +0800)]
[SYCL] Fix building Win package for oneAPI 2025.0 update (#10483)
* fix build package for 2025.0
* debug
* debug
* fix
* rm debug
---------
Co-authored-by: arthw <redacted>
Georgi Gerganov [Mon, 25 Nov 2024 07:58:41 +0000 (09:58 +0200)]
speculative : refactor and add a simpler example (#10362)
* speculative : refactor and add a simpler example
ggml-ci
* speculative : clean-up and add comments and TODOs [no ci]
* speculative : manage context in common_speculative
ggml-ci
* speculative : simplify
ggml-ci
* speculative : simplify (cont)
ggml-ci
* speculative : add --draft-min CLI arg
* speculative : minor fixup
* make : build fixes
* speculative : do not redraft previous drafts
ggml-ci
* speculative : fix the draft sampling
ggml-ci
* speculative : fix compile warning
* common : refactor args
ggml-ci
* common : change defaults [no ci]
* common : final touches
ggml-ci
Georgi Gerganov [Sun, 24 Nov 2024 16:03:25 +0000 (18:03 +0200)]
flake.lock: Update (#10470)
Flake lock file updates:
• Updated input 'nixpkgs':
'github:NixOS/nixpkgs/
5e4fbfb6b3de1aa2872b76d49fafc942626e2add ?narHash=sha256-OZiZ3m8SCMfh3B6bfGC/Bm4x3qc1m2SVEAlkV6iY7Yg%3D' (2024-11-15)
→ 'github:NixOS/nixpkgs/
23e89b7da85c3640bbc2173fe04f4bd114342367 ?narHash=sha256-y/MEyuJ5oBWrWAic/14LaIr/u5E0wRVzyYsouYY3W6w%3D' (2024-11-19)
Co-authored-by: github-actions[bot] <redacted>
Diego Devesa [Sun, 24 Nov 2024 15:10:26 +0000 (16:10 +0100)]
llama : fix op mul check with command-r-plus (#10476)
Gabe Goodhart [Sun, 24 Nov 2024 09:02:34 +0000 (02:02 -0700)]
convert : XLMRoberta Type Vocab Size (#10458)
This matches the key in common bert-based embedding models and may have a
value other than 1 in it.
Branch: XLMRobertaTypeVocabSize
Signed-off-by: Gabe Goodhart <redacted>
momonga [Sun, 24 Nov 2024 00:09:22 +0000 (09:09 +0900)]
fix gguf-py: Conversion error when multiple licenses are configured (#9807)
* fix general.license list to str
* fix join license list
---------
Co-authored-by: momonga <redacted>
Diego Devesa [Sat, 23 Nov 2024 13:41:12 +0000 (14:41 +0100)]
ggml : do not use ARM features not included in the build (#10457)
蕭澧邦 [Fri, 22 Nov 2024 09:44:08 +0000 (17:44 +0800)]
ci: Update oneAPI runtime dll packaging (#10428)
This is the minimum runtime dll dependencies for oneAPI 2025.0
Johannes Gäßler [Fri, 22 Nov 2024 07:32:40 +0000 (08:32 +0100)]
GitHub: ask for more info in issue templates (#10426)
* GitHub: ask for more info in issues [no ci]
* refactor issue templates to be component-specific
* more understandable issue description
* add dropdown for llama.cpp module
leo-pony [Fri, 22 Nov 2024 06:07:20 +0000 (14:07 +0800)]
CANN: Support Ascend310P to accelerate F32 and F16 Model (#10216)
* CANN Support Ascend310P to accelerate F32 and F16 Model
* Add compile option soc type macro ASCEND_310P to ggml-cann lib
* Remove unused code
* Remove the ascend soc_type hard code compile option in CMakelist.txt
Diego Devesa [Thu, 21 Nov 2024 17:18:50 +0000 (18:18 +0100)]
cuda : optimize argmax (#10441)
* cuda : optimize argmax
* remove unused parameter
ggml-ci
* fixup : use full warps
ggml-ci
* Apply suggestions from code review
Co-authored-by: Johannes Gäßler <redacted>
* fix ub
* ggml : check ne00 <= INT32_MAX in argmax and argsort
---------
Co-authored-by: Johannes Gäßler <redacted>
Georgi Gerganov [Thu, 21 Nov 2024 08:22:47 +0000 (10:22 +0200)]
llama : handle KV shift for recurrent models (#10402)
ggml-ci
Georgi Gerganov [Thu, 21 Nov 2024 07:22:11 +0000 (09:22 +0200)]
sync : ggml
slaren [Wed, 20 Nov 2024 12:25:08 +0000 (13:25 +0100)]
ggml/sched : do not skip views in pre-assignments
Johannes Gäßler [Wed, 20 Nov 2024 13:56:04 +0000 (14:56 +0100)]
ggml-opt: fix data corruption (ggml/1022)
Jeff Bolz [Wed, 20 Nov 2024 19:47:36 +0000 (13:47 -0600)]
vulkan: predicate max operation in soft_max shaders/soft_max (#10437)
Fixes #10434
bandoti [Wed, 20 Nov 2024 16:22:19 +0000 (12:22 -0400)]
cmake: add link dependencies to cmake find pkg (#10433)
* cmake pkg: find accelerate, openmp, memkind libs
* cmake pkg: find BLAS libs
* try BLAS_LIBRARIES instead
* Add BLAS link opts
* Add more link deps. and set GGML_ vars
Diego Devesa [Wed, 20 Nov 2024 11:57:53 +0000 (12:57 +0100)]
llama : add .clang-format file (#10415)
Jeff Bolz [Wed, 20 Nov 2024 07:40:18 +0000 (01:40 -0600)]
vulkan: copy iq4_nl LUT into shared memory (#10409)
Jeff Bolz [Wed, 20 Nov 2024 07:11:00 +0000 (01:11 -0600)]
vulkan: further optimize mul_mat_vec using larger loads (#10387)
* vulkan: Use pipeline_robustness to disable robustness in mul_mat_vec.
Add some early returns for nonexistent rows in mul_mat_vec shaders. These
can only be hit when dispatching a 2D grid of workgroups. Fix the logic
for the 2D grid of workgroups to round up.
Enable the pipeline robustness extension if it's available, and use it to
disable robustness for these pipelines. The instructions to do the bounds
checking contend for the same ALU resources as the bit twiddling dequant
instructions.
* vulkan: Add GLSL structure aliases for quant types to allow larger loads
In Vulkan it's not possible to cast pointer types, so instead you have to
declare an aliased binding for the memory with a different type. This
commit adds aliases for the quant formats using 16b ints, and in a few
places where the struct size is a multiple of 4 also using 32b ints.
Currently only q4_k's aliases are used, but others will be used in
subsequent commits.
* vulkan: use larger loads in q5_k and q6_k shaders.
Similar to the optimization I did in q4_k recently, this vectorizes some loads
and reduces the number of bit twiddling instructions.
* vulkan: use larger K step per iteration in mul_mat_vec.
Add vec4 dequantization functions, and use them to do K=8 per iteration in
mul_mat_vec. This uses 16b loads for the quant values and 128b loads for B
which helps reduce the load on the memory system.
The K_PER_ITER==2 logic is still there, just for F16/F32, and really only
because they support unaligned sizes.
Tweak the num_iters/unrolling logic to be simpler and catch a couple missed
unrolling opportunities.
Neo Zhang Jianyu [Wed, 20 Nov 2024 05:54:25 +0000 (13:54 +0800)]
update rel to 4040 (#10395)
Co-authored-by: arthw <redacted>
Anthony Van de Gejuchte [Tue, 19 Nov 2024 22:18:17 +0000 (23:18 +0100)]
Fix missing file renames in Makefile due to changes in commit
ae8de6d50a (#10413)
haopeng [Tue, 19 Nov 2024 20:10:31 +0000 (04:10 +0800)]
add cmake rvv support (#10411)
Georgi Gerganov [Tue, 19 Nov 2024 17:15:50 +0000 (19:15 +0200)]
sync : ggml
Plamen Minev [Mon, 18 Nov 2024 13:02:27 +0000 (15:02 +0200)]
metal : fox offset integer overflows in im2col (ggml/1015)
-- While running StableDiffusion.cpp locally with Metal some offsets overflow and results in incorrect calculations
PAB [Mon, 18 Nov 2024 09:02:49 +0000 (10:02 +0100)]
metal : add `GGML_UNARY_OP_ELU` kernel (ggml/1018)
蕭澧邦 [Tue, 19 Nov 2024 17:42:00 +0000 (01:42 +0800)]
cmake: force MSVC compiler charset to utf-8 (#9989)
bandoti [Tue, 19 Nov 2024 16:10:30 +0000 (12:10 -0400)]
Add required ggml-base and backend libs to cmake pkg (#10407)
Diego Devesa [Tue, 19 Nov 2024 13:29:38 +0000 (14:29 +0100)]
cuda : fix CUDA_FLAGS not being applied (#10403)
Georgi Gerganov [Tue, 19 Nov 2024 11:29:26 +0000 (13:29 +0200)]
llama : add check for KV cache shifts (#10401)
ggml-ci
Shane A [Tue, 19 Nov 2024 09:04:08 +0000 (01:04 -0800)]
llama : add OLMo November 2024 support (#10394)
* Add OLMo November 2024 constants
* Add OLMo November 2024 converter
* Add loading of OLMo November 2024 tensors and hyper parameters
* Add building of OLMo November 2024 model
Romain Biessy [Tue, 19 Nov 2024 08:02:23 +0000 (09:02 +0100)]
sycl : Add option to set the SYCL architecture for all targets (#10266)
* Add option to set the SYCL architecture for all targets
* Convert GGML_SYCL_HIP_TARGET to the more generic GGML_SYCL_ARCH option
* Document that setting GGML_SYCL_ARCH can improve the performance
Jeff Bolz [Tue, 19 Nov 2024 07:25:17 +0000 (01:25 -0600)]
vulkan: Optimize soft_max (#10301)
* vulkan: Optimize soft_max
Large soft_max could already saturate memory, but small/medium sizes were
pretty slow. The bulk of the gains for them comes from using a smaller
workgroup size, and making the workgroup size match the subgroup size also
makes the barriers much cheaper.
Cache some values in locals to avoid refetching/recomputing. And stamp
out a few "template instantiations" so smaller cases will fully unroll.
Add a missing early return for OOB rows. This happens when there are more
than 512 rows and the dispatch is 512 x H.
* vulkan: Further soft_max optimizations
Restore the workgroup size of 512 case, use it for >1024.
Use unrollable loops for more iteration counts.
Alberto Cabrera Pérez [Tue, 19 Nov 2024 00:50:04 +0000 (00:50 +0000)]
sycl: Revert MUL_MAT_OP support changes (#10385)
Diego Devesa [Mon, 18 Nov 2024 17:43:40 +0000 (18:43 +0100)]
cuda : only use native when supported by cmake (#10389)
bandoti [Mon, 18 Nov 2024 15:23:58 +0000 (11:23 -0400)]
Skip searching root path for cross-compile builds (#10383)
Jeff Bolz [Mon, 18 Nov 2024 14:28:42 +0000 (08:28 -0600)]
vulkan: remove use of null initializer (#10372)
Seems like this isn't working for vulkan-over-metal when the array is sized
by a spec constant. Maybe a spirv-cross limitation?
Georgi Gerganov [Mon, 18 Nov 2024 14:08:20 +0000 (16:08 +0200)]
flake.lock: Update (#10346)
Flake lock file updates:
• Updated input 'nixpkgs':
'github:NixOS/nixpkgs/
4aa36568d413aca0ea84a1684d2d46f55dbabad7 ?narHash=sha256-Zwl8YgTVJTEum%2BL%2B0zVAWvXAGbWAuXHax3KzuejaDyo%3D' (2024-11-05)
→ 'github:NixOS/nixpkgs/
5e4fbfb6b3de1aa2872b76d49fafc942626e2add ?narHash=sha256-OZiZ3m8SCMfh3B6bfGC/Bm4x3qc1m2SVEAlkV6iY7Yg%3D' (2024-11-15)
Co-authored-by: github-actions[bot] <redacted>
0cc4m [Mon, 18 Nov 2024 10:02:43 +0000 (11:02 +0100)]
Vulkan: Fix device info output format specifiers (#10366)
* Vulkan: Fix device info output format specifiers
* Vulkan: Use zu printf specifier for size_t instead of ld
Johannes Gäßler [Sun, 17 Nov 2024 23:21:53 +0000 (00:21 +0100)]
docker: use GGML_NATIVE=OFF (#10368)
Johannes Gäßler [Sun, 17 Nov 2024 22:20:42 +0000 (23:20 +0100)]
CUDA: fix MMV kernel being used for FP16 src1 (#10357)
Johannes Gäßler [Sun, 17 Nov 2024 11:59:38 +0000 (12:59 +0100)]
CMake: fix typo in comment [no ci] (#10360)
Diego Devesa [Sun, 17 Nov 2024 11:25:45 +0000 (12:25 +0100)]
llama : only use default buffer types for the KV cache (#10358)
Georgi Gerganov [Sun, 17 Nov 2024 11:12:22 +0000 (13:12 +0200)]
gitignore : ignore local run scripts [no ci]
Georgi Gerganov [Sun, 17 Nov 2024 09:23:01 +0000 (11:23 +0200)]
metal : refactor kernel args into structs (#10238)
* metal : add kernel arg structs (wip)
* metal : fattn args
ggml-ci
* metal : cont + avoid potential int overflow [no ci]
* metal : mul mat struct (wip)
* cont : mul mat vec
* cont : pass by reference
* cont : args is first argument
* cont : use char ptr
* cont : shmem style
* cont : thread counters style
* cont : mul mm id
ggml-ci
* cont : int safety + register optimizations
ggml-ci
* metal : GGML_OP_CONCAT
ggml-ci
* metal : GGML_OP_ADD, GGML_OP_SUB, GGML_OP_MUL, GGML_OP_DIV
* metal : GGML_OP_REPEAT
* metal : GGML_OP_CPY
* metal : GGML_OP_RMS_NORM
* metal : GGML_OP_NORM
* metal : add TODOs for rest of ops
* ggml : add ggml-metal-impl.h
ggml-ci
FirstTimeEZ [Sun, 17 Nov 2024 08:39:22 +0000 (21:39 +1300)]
ggml : fix undefined reference to 'getcpu' (#10354)
https://github.com/ggerganov/llama.cpp/issues/10352
Johannes Gäßler [Sun, 17 Nov 2024 08:09:55 +0000 (09:09 +0100)]
CUDA: remove DMMV, consolidate F16 mult mat vec (#10318)
Johannes Gäßler [Sun, 17 Nov 2024 08:06:34 +0000 (09:06 +0100)]
CMake: default to -arch=native for CUDA build (#10320)
Diego Devesa [Sun, 17 Nov 2024 06:31:17 +0000 (07:31 +0100)]
ggml : fix possible buffer use after free in sched reserve (#9930)
Georgi Gerganov [Sat, 16 Nov 2024 21:40:39 +0000 (23:40 +0200)]
ggml : inttypes.h -> cinttypes (#0)
ggml-ci
Georgi Gerganov [Sat, 16 Nov 2024 19:38:01 +0000 (21:38 +0200)]
ggml : adapt AMX to tensor->grad removal (#0)
ggml-ci
Georgi Gerganov [Sat, 16 Nov 2024 19:35:31 +0000 (21:35 +0200)]
make : add ggml-opt (#0)
ggml-ci
Georgi Gerganov [Sat, 16 Nov 2024 19:34:03 +0000 (21:34 +0200)]
tests : remove test-grad0
Georgi Gerganov [Sat, 16 Nov 2024 19:32:41 +0000 (21:32 +0200)]
ggml : fix compile warnings (#0)
ggml-ci
Johannes Gäßler [Sat, 16 Nov 2024 20:17:59 +0000 (22:17 +0200)]
ggml: new optimization interface (ggml/988)
Georgi Gerganov [Sat, 16 Nov 2024 20:16:04 +0000 (22:16 +0200)]
scripts : update sync
FirstTimeEZ [Sat, 16 Nov 2024 23:29:18 +0000 (12:29 +1300)]
docs : vulkan build instructions to use git bash mingw64 (#10303)
Johannes Gäßler [Sat, 16 Nov 2024 22:00:41 +0000 (23:00 +0100)]
llama/ex: remove --logdir argument (#10339)
Georgi Gerganov [Sat, 16 Nov 2024 15:58:56 +0000 (17:58 +0200)]
llamafile : fix include path (#0)
ggml-ci
Georgi Gerganov [Sat, 16 Nov 2024 15:58:32 +0000 (17:58 +0200)]
make : auto-determine dependencies (#0)
MaggotHATE [Sat, 16 Nov 2024 13:26:54 +0000 (18:26 +0500)]
server: (web UI) Add samplers sequence customization (#10255)
* Samplers sequence: simplified and input field.
* Removed unused function
* Modify and use `settings-modal-short-input`
* rename "name" --> "label"
---------
Co-authored-by: Xuan Son Nguyen <redacted>
Georgi Gerganov [Sat, 16 Nov 2024 08:32:50 +0000 (10:32 +0200)]
scripts : fix missing key in compare-llama-bench.py (#10332)
Jeff Bolz [Sat, 16 Nov 2024 06:26:57 +0000 (00:26 -0600)]
vulkan: Optimize some mat-vec mul quant shaders (#10296)
Compute two result elements per workgroup (for Q{4,5}_{0,1}). This reuses
the B loads across the rows and also reuses some addressing calculations.
This required manually partially unrolling the loop, since the compiler
is less willing to unroll outer loops.
Add bounds-checking on the last iteration of the loop. I think this was at
least partly broken before.
Optimize the Q4_K shader to vectorize most loads and reduce the number of
bit twiddling instructions.
FirstTimeEZ [Sat, 16 Nov 2024 01:59:33 +0000 (14:59 +1300)]
vulkan : add cmake preset debug/release (#10306)
Dan Johansson [Sat, 16 Nov 2024 00:53:37 +0000 (01:53 +0100)]
ggml : optimize Q4_0 into Q4_0_X_Y repack (#10324)
FirstTimeEZ [Sat, 16 Nov 2024 00:42:13 +0000 (13:42 +1300)]
llama : save number of parameters and the size in llama_model (#10286)
fixes #10285
Srihari-mcw [Fri, 15 Nov 2024 21:27:00 +0000 (02:57 +0530)]
Make updates to fix issues with clang-cl builds while using AVX512 flags (#10314)
Johannes Gäßler [Fri, 15 Nov 2024 20:19:03 +0000 (21:19 +0100)]
scripts: update compare-llama-bench.py (#10319)
slaren [Fri, 15 Nov 2024 19:20:54 +0000 (20:20 +0100)]
ggml : fix some build issues
Georgi Gerganov [Fri, 15 Nov 2024 13:35:22 +0000 (15:35 +0200)]
cmake : fix ppc64 check (whisper/0)
ggml-ci
thewh1teagle [Fri, 15 Nov 2024 13:33:53 +0000 (15:33 +0200)]
ggml : vulkan logs (whisper/2547)
Georgi Gerganov [Fri, 15 Nov 2024 13:31:16 +0000 (15:31 +0200)]
sync : ggml
Eve [Fri, 15 Nov 2024 11:47:58 +0000 (11:47 +0000)]
AVX BF16 and single scale quant optimizations (#10212)
* use 128 bit loads (i've tried 256->128 to death and its slower)
* double accumulator
* avx bf16 vec dot
* +3% q4_0 inference
* +7% tg +5% pp compared to master
* slower f16c version, kep for reference
* 256b version, also slow. i tried :)
* revert f16
* faster with madd
* split to functions
* Q8_0 and IQ4_NL, 5-7% faster
* fix potential overflow (performance reduced)
* 16 bit add for q4_0 only
* merge
R0CKSTAR [Fri, 15 Nov 2024 11:47:25 +0000 (19:47 +0800)]
ci: build test musa with cmake (#10298)
Signed-off-by: Xiaodong Ye <redacted>
Romain Biessy [Fri, 15 Nov 2024 11:10:45 +0000 (12:10 +0100)]
sycl: Update Intel docker images to use DPC++ 2025.0 (#10305)
Xuan Son Nguyen [Fri, 15 Nov 2024 09:48:49 +0000 (05:48 -0400)]
server : (web UI) add copy button for code block, fix api key (#10242)
* server : (web ui) add copy btn for code blocks
* fix problem with api key
* use settings-modal-short-input component
* always show copy btn for code snippet
Chenguang Li [Fri, 15 Nov 2024 07:09:35 +0000 (15:09 +0800)]
cann: dockerfile and doc adjustment (#10302)
Co-authored-by: noemotiovon <redacted>
Georgi Gerganov [Fri, 15 Nov 2024 06:38:43 +0000 (08:38 +0200)]
scripts : fix regex in sync [no ci]
Romain Biessy [Fri, 15 Nov 2024 03:09:12 +0000 (04:09 +0100)]
sycl: Use syclcompat::dp4a (#10267)
* sycl: Use syclcompat::dp4a
* Using the syclcompat version allow the compiler to optimize the
operation with native function
* Update news section
* Update CI Windows oneAPI version to 2025.0
* Reword doc
* Call syclcompat::dp4a inside dpct::dp4a
This reverts commit
90cb61d692d61360b46954a1c7f780bd2e569b73 .
Charles Xu [Fri, 15 Nov 2024 00:28:50 +0000 (01:28 +0100)]
backend cpu: add online flow for aarch64 Q4_0 GEMV/GEMM kernels (#9921)
* backend-cpu: add online flow for aarch64 Q4_0 GEMV/GEMM kernels
---------
Co-authored-by: Diego Devesa <redacted>
Diego Devesa [Thu, 14 Nov 2024 17:04:35 +0000 (18:04 +0100)]
ggml : build backends as libraries (#10256)
* ggml : build backends as libraries
---------
Signed-off-by: Xiaodong Ye <redacted>
Co-authored-by: Georgi Gerganov <redacted>
Co-authored-by: R0CKSTAR <redacted>
Johannes Gäßler [Thu, 14 Nov 2024 12:00:15 +0000 (13:00 +0100)]
CUDA: no -sm row for very small matrices (#10185)
Georgi Gerganov [Thu, 14 Nov 2024 09:44:15 +0000 (11:44 +0200)]
speculative : fix out-of-bounds access (#10289)
Jeff Bolz [Thu, 14 Nov 2024 05:22:55 +0000 (23:22 -0600)]
vulkan: Optimize binary ops (#10270)
Reuse the index calculations across all of src0/src1/dst. Add a shader
variant for when src0/src1 are the same dimensions and additional modulus
for src1 aren't needed. Div/mod are slow, so add "fast" div/mod that
have a fast path when the calculation isn't needed or can be done more
cheaply.
Jeff Bolz [Wed, 13 Nov 2024 20:59:47 +0000 (14:59 -0600)]
vulkan: Use macros to make the mat mul pipeline creation more concise (#10259)
Also add vk_matmul_pipeline2 to hold f16/f32 accumulator versions of a
pipeline. This isn't really used yet.