33d7aed4a8
CANN: Support MOE Model MUL_MAT_ID ( #13042 )
...
Signed-off-by: noemotiovon <757486878@qq.com >
2025-05-19 14:21:17 +08:00
e3a7cf6c5b
cmake: use the current build config for vulkan-shaders-gen ( #13595 )
...
* fix: use the current build config for `vulkan-shaders-gen`
* fix: only pass a valid build type to `--config`
2025-05-17 15:26:43 -03:00
2f5a4e1e09
vulkan: move common FA code to flash_attn_base.comp ( #13556 )
...
* vulkan: move common FA code to flash_attn_base.comp
* vulkan: move common FA index/stride setup code to flash_attn_base.comp
* build fix
2025-05-17 09:14:55 +02:00
4f41ee11d6
vulkan: use scalar FA rather than coopmat2 when N==1 ( #13554 )
2025-05-17 08:35:47 +02:00
654a67794f
metal : add FA-vec kernel for head size 64 ( #13583 )
...
ggml-ci
2025-05-16 20:32:58 +03:00
0a338ed013
sycl : fixed compilation warnings ( #13582 )
2025-05-16 18:15:29 +08:00
c6a2c9e741
gguf : use ggml log system ( #13571 )
...
* gguf : use ggml log system
* llama : remove unnecessary new lines in exception messages
2025-05-15 19:13:11 +02:00
02cdd2d8b0
sycl: simplify bin_bcast_kernel ( #13383 )
2025-05-15 17:39:52 +02:00
64bb51cf90
sycl: reordered Q4_K MMVQ ( #13109 )
2025-05-15 17:35:44 +02:00
9c404ed54c
sycl: use oneDNN for matrices multiplication ( #12972 )
2025-05-15 16:53:41 +02:00
5ab5d5fb25
arm64: optimize q6_k_q8_k kernel with i8mm ( #13519 )
...
This PR improves q6_k_q8_k gemm kernel with arm64 i8mm instruction.
Tested on neoverse-n2 with llama3 8b q6_k quantization model.
- 40% ~ 54% S_PP uplift for all batch sizes
- 16% ~ 47% S_TG uplift for batch size 4 and above
Perplexity doesn't change with this PR.
```
// tested on neoverse-n2
$ llama-batched-bench \
-m Meta-Llama-3-8B-Instruct-Q6_K.gguf \
--no-mmap -fa \
-c 8192 -b 4096 -ub 512 -npp 128 -ntg 128 \
-npl 1,2,4,8,16,32 \
-t 64
---------------------------------------------------------------------
| PP | TG | B | S_PP t/s | S_TG t/s |
| | | | original | this pr | original | this pr |
|-------|--------|------|----------|----------|----------|----------|
| 128 | 128 | 1 | 78.52 | 109.18 | 18.63 | 18.88 |
| 128 | 128 | 2 | 84.62 | 123.94 | 34.54 | 36.92 |
| 128 | 128 | 4 | 84.36 | 122.49 | 52.65 | 61.32 |
| 128 | 128 | 8 | 90.52 | 138.87 | 63.46 | 84.41 |
| 128 | 128 | 16 | 90.11 | 138.56 | 71.04 | 101.33 |
| 128 | 128 | 32 | 89.81 | 137.79 | 75.14 | 110.47 |
---------------------------------------------------------------------
```
2025-05-14 21:53:52 +02:00
4696d56749
CUDA: fix crash on large batch size for quant. MoE ( #13537 )
2025-05-14 16:41:02 +02:00
6da34fa276
CUDA: faster Deepseek FA, add Turing support ( #13435 )
2025-05-14 16:08:20 +02:00
09d13d94fb
cmake: simplify vulkan shader test logic ( #13263 )
2025-05-14 07:53:57 -03:00
24e86cae72
vulkan: KHR_coopmat flash attention ( #13506 )
...
This shader uses coopmat1 to do the Q*K^T multiply. The P*V multiply is more
difficult for various reasons so I haven't done it. Performance for this
shader is around 2.5x better than for the scalar shader when doing prompt
processing. Some of the benefit may be from other optimizations like staging
through shared memory, or splitting by rows.
2025-05-14 11:55:26 +02:00
ab3971f2a0
vulkan: workaround FA compile failures on macos ( #13517 )
2025-05-14 06:15:50 +02:00
f0995d28ce
metal : use FA-vec kernel up to batch size 20 ( #13496 )
...
* batched-bench : fix pp batch contents
* metal : optimize multi-sequence FA vec kernel
ggml-ci
* metal : use FA-vec kernel up to batch size 20
ggml-ci
2025-05-13 18:04:39 +03:00
c252e0c409
metal : optimize multi-sequence FA vec kernel ( #13493 )
...
* batched-bench : fix pp batch contents
* metal : optimize multi-sequence FA vec kernel
ggml-ci
2025-05-13 18:04:00 +03:00
4f711afed5
ggml-cpu: Update KleidiAI to v1.6 and fix include directives ( #13509 )
...
Signed-off-by: Dan Johansson <dan.johansson@arm.com >
2025-05-13 18:02:28 +03:00
f0d46ef157
opencl: remove unnecessary assert for add ( #13257 )
2025-05-12 13:13:49 -07:00
10d2af0eaa
llama/ggml: add LLM training support ( #10544 )
...
* llama/ggml: add LLM training support
more compact progress bar
llama_save_model_to_file
llama_opt_param_filter
ggml_graph_dup force_grads
refactor ggml_opt, fix test-opt
* remove logits_all
* refactor CUDA implementation for ACC
* reset graph at beginning of opt period
2025-05-12 14:44:49 +02:00
a71a4075cd
ggml-cpu: Integrate fp32=bf16xbf16 SME KleidiAI kernel ( #13053 )
...
* ggml-cpu: Integrate fp32=bf16xbf16 SME KleidiAI kernel
Signed-off-by: Dan Johansson <dan.johansson@arm.com >
* * code review fixes
Signed-off-by: Dan Johansson <dan.johansson@arm.com >
* * adds a comment that clarifies barrier usage
Signed-off-by: Dan Johansson <dan.johansson@arm.com >
---------
Signed-off-by: Dan Johansson <dan.johansson@arm.com >
Co-authored-by: Charles Xu <charles.xu@arm.com >
2025-05-12 13:06:19 +02:00
95e18884fc
CUDA: fix misaligned synchronization in FA ( #13469 )
2025-05-12 10:51:21 +02:00
df8491922f
ggml : add mrope kernel for metal ( #13457 )
2025-05-12 10:29:13 +02:00
14492144c2
enable dpcpp nightly builds with libraries ( #13406 )
2025-05-12 13:15:32 +08:00
7474e00b34
CUDA: fix crash with partial offloading of MoE ( #13439 )
2025-05-11 16:09:33 +02:00
7f323a589f
Add --no-op-offload to improve -ot pp perf in MoE models like llama4 400B ( #13386 )
2025-05-11 14:18:39 +02:00
0208355f42
CUDA: fix race conditions FlashAttention kernels ( #13438 )
2025-05-10 22:22:48 +02:00
d8919424f1
CUDA: fix FlashAttention on Turing ( #13415 )
2025-05-10 09:16:52 +02:00
dc1d2adfc0
vulkan: scalar flash attention implementation ( #13324 )
...
* vulkan: scalar flash attention implementation
* vulkan: always use fp32 for scalar flash attention
* vulkan: use vector loads in scalar flash attention shader
* vulkan: remove PV matrix, helps with register usage
* vulkan: reduce register usage in scalar FA, but perf may be slightly worse
* vulkan: load each Q value once. optimize O reduction. more tuning
* vulkan: support q4_0/q8_0 KV in scalar FA
* CI: increase timeout to accommodate newly-supported tests
* vulkan: for scalar FA, select between 1 and 8 rows
* vulkan: avoid using Float16 capability in scalar FA
2025-05-10 08:07:07 +02:00
17512a94d6
sycl : implementation of reordered Q4_0 MMVQ for Intel GPUs ( #12858 )
...
* sycl : Implemented reorder Q4_0 mmvq
Signed-off-by: Alberto Cabrera <alberto.cabrera@codeplay.com >
* sycl : Fixed mmvq being called when reorder is disabled
* sycl : Improved comments in the quants header
Signed-off-by: Alberto Cabrera <alberto.cabrera@codeplay.com >
* Use static_assert
* safe_div -> ceil_div
* Clarify qi comment
* change the reorder tensor from init to execute OP
* dbg
* Undo changes to test-backend-ops
* Refactor changes on top of q4_0 reorder fix
* Missing Reverts
* Refactored opt_for_reorder logic to simplify code path
* Explicit inlining and unroll
* Renamed mul_mat_algo enum for consistency
---------
Signed-off-by: Alberto Cabrera <alberto.cabrera@codeplay.com >
Co-authored-by: romain.biessy <romain.biessy@codeplay.com >
2025-05-09 16:34:08 +01:00
611aa914ef
metal : optimize MoE for large batches ( #13388 )
...
ggml-ci
2025-05-09 15:14:56 +03:00
0cf6725e9f
CUDA: FA support for Deepseek (Ampere or newer) ( #13306 )
...
* CUDA: FA support for Deepseek (Ampere or newer)
* do loop unrolling via C++ template
2025-05-09 13:34:58 +02:00
5c86c9ed3e
CUDA: fix crash on large batch size for MoE models ( #13384 )
2025-05-09 12:14:04 +02:00
b486ba05bf
rpc : add rpc_msg_set_tensor_hash_req ( #13353 )
...
* rpc : add rpc_msg_set_tensor_hash_req
Use a dedicated struct for the request of RPC_CMD_SET_TENSOR_HASH which
makes the code cleaner.
* fix
2025-05-09 10:31:07 +03:00
02115dcd9a
vulkan: Allow up to 4096 elements for mul_mat_id row_ids ( #13326 )
...
This assert fired running Qwen_Qwen3-30B-A3B-Q2_K.gguf:
GGML_ASSERT(nei0 * nei1 <= 3072);
The tensor is 8 x 512. Increase this array size to accommodate.
2025-05-09 09:23:41 +02:00
8733e0cf6e
sycl: addressing non-contiguous src1 mul_mats (nc and batched) ( #13343 )
...
* sycl: fixed non-contiguous src1 mul_mats (nc and batched)
* Fixed wrong static_cast inside kernel
2025-05-08 10:08:01 +01:00
13b0a04597
whisper: remove MSVC warnings pragmas (whisper/3090)
...
* ggml : remove MSVC warnings pragmas
This commit removes the MSVC-specific pragmas as these are now handled
in ggml/CMakeLists.txt.
* whisper : remove MSVC warning pragmas
This commit removes the MSVC-specific pragmas. These are now handled in
the ggml/CMakeLists.txt file.
2025-05-07 17:28:36 +03:00
bba9d945c1
cmake : removed stdc++fs (whisper/3097)
...
* removed stdc++fs
* kept line, but removed stdc++fs
2025-05-07 17:28:36 +03:00
1f73301b63
cuda : remove nrows_x in mul_mat_q_process_tile ( #13325 )
...
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com >
2025-05-07 09:48:23 +02:00
141a908a59
CUDA: mix virt/real CUDA archs for GGML_NATIVE=OFF ( #13135 )
2025-05-06 23:35:51 +02:00
1e333d5bba
SYCL: Disable reorder optimize by default and stop setting tensor extras when optimize is disabled ( #13254 )
...
* SYCL: Do not set tensor extras when reorder optimize is disabled
* SYCL: Disable reorder optimize by default
2025-05-06 20:27:06 +05:30
2356fb1d53
CUDA: fix bad asserts for partial offload ( #13337 )
2025-05-06 13:58:51 +02:00
15a28ec8c7
CUDA: fix --split-mode row for MMQ ( #13323 )
2025-05-06 08:36:46 +02:00
9070365020
CUDA: fix logic for clearing padding with -ngl 0 ( #13320 )
2025-05-05 22:32:13 +02:00
66645a5285
SYCL: Disable mul_mat kernels for noncontiguous tensor b ( #13308 )
...
ggml-ci
2025-05-05 13:39:10 +05:30
9fdfcdaedd
rpc : use backend registry, support dl backends ( #13304 )
2025-05-04 21:25:43 +02:00
6eb7d25c70
ggml : activate s390x simd for Q3_K ( #13301 )
...
Signed-off-by: Aaron Teo <aaron.teo1@ibm.com >
2025-05-04 19:49:12 +02:00
93c4e23905
CUDA: fix race condition in MMQ stream-k fixup ( #13299 )
2025-05-04 14:16:39 +02:00
8afbd96818
CUDA: fix race condition in MMQ ids_dst ( #13294 )
2025-05-04 13:58:38 +02:00
Chenguang Li