mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 31 Packages Projects Releases Wiki Activity
Files
ecbcb7ea9d3303097519723b264a8b5f1e977028
llama.cpp-mtp-turboquant/ggml/src/ggml-cuda
T
History
Jayant Lohia ecbcb7ea9d CUDA: add CDNA3 MFMA support for flash attention MMA kernel (#19806) 
* CUDA: add CDNA3 MFMA support for flash attention MMA kernel

Add MI300X (gfx942) MFMA tensor core flash attention using
v_mfma_f32_16x16x16_f16 (FP16 in, FP32 accumulate).

- Add FATTN_WARP_SIZE=64 for CDNA wavefront64
- Add CDNA config for head sizes 64, 80, 96, 112, 128
- Add FP16 MFMA intrinsic path in mma.cuh
- Add manual V transpose load for MFMA register layout
- Route CDNA to MMA for prompt processing, VEC for token generation
- Fix Q loading and combine stride granularity for non-power-of-2 heads

Benchmarks (Qwen2.5-1.5B Q4_K_M, MI300X):
  pp512  +7%,  pp1024 +13%,  pp2048 +23%,  pp4096 +39%
  tg128  -10% (FA overhead, VEC used for both)

All 2480 flash attention tests pass.

Ref: https://github.com/ggml-org/llama.cpp/issues/17917

* address review: replace FATTN_WARP_SIZE with constexpr, improve dispatch

- Replace #define FATTN_WARP_SIZE with constexpr int warp_size =
  ggml_cuda_get_physical_warp_size() in each device function
- Use ne[1]*gqa_ratio threshold for MMA vs tile dispatch. Benchmarked
  crossover on MI300X @ d32768 with power-of-2 GQA models:
    hsk=64  (Llama 1B, gqa=4): MMA wins at eff >= 128 (+11%)
    hsk=128 (Llama 3B, gqa=4): MMA wins at eff >= 128 (+4%)
  Unified threshold: eff_nq >= 128 for all head sizes.
- Remove VEC fallback; small batches fall through to tile kernel

* Update ggml/src/ggml-cuda/fattn.cu

* use ggml_cuda_info().devices warp_size instead of hardcoded check

---------

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2026-02-27 19:37:26 +01:00
..
template-instances
CUDA: faster FA for GQA > 1 but not power of 2 (#19092)
2026-01-25 21:19:47 +01:00
vendors
HIP: add fattn-mma-f16 for RDNA4 (#18481)
2026-01-13 13:52:16 +01:00
acc.cu
llama/ggml: add LLM training support (#10544)
2025-05-12 14:44:49 +02:00
acc.cuh
llama : reorganize source code + improve CMake (#8006)
2024-06-26 18:33:02 +03:00
add-id.cu
musa: fix build warnings (#15258)
2025-08-20 10:17:37 +08:00
add-id.cuh
llama : add gpt-oss (#15091)
2025-08-05 22:10:36 +03:00
arange.cu
llama : reorganize source code + improve CMake (#8006)
2024-06-26 18:33:02 +03:00
arange.cuh
llama : reorganize source code + improve CMake (#8006)
2024-06-26 18:33:02 +03:00
argmax.cu
ggml : use WARP_SIZE/2 for argmax reduction offset (#18092)
2025-12-17 11:47:01 +08:00
argmax.cuh
ggml/ex: calculate accuracy in graph, adapt MNIST (ggml/980)
2024-10-03 21:17:26 +03:00
argsort.cu
CUDA: Fix builds for older CCCL versions by ifdefing strided_iterator (#18964)
2026-01-21 02:34:29 +01:00
argsort.cuh
sampling : add support for backend sampling (#17004)
2026-01-04 22:22:16 +02:00
binbcast.cu
ggml : extend bin bcast for permuted src1 (#19484)
2026-02-11 07:52:00 +02:00
binbcast.cuh
CUDA: fuse adds, fuse add with rms norm (#15631)
2025-08-29 11:35:58 +08:00
clamp.cu
cuda: unary ops as float + de-duplicate (ggml/1130)
2025-03-03 18:18:11 +02:00
clamp.cuh
llama : reorganize source code + improve CMake (#8006)
2024-06-26 18:33:02 +03:00
CMakeLists.txt
CUDA : Update CCCL-tag for 3.2 to final release from RC (#19486)
2026-02-10 22:31:19 +01:00
common.cuh
Improve CUDA graph capture (#19754)
2026-02-21 15:09:36 +05:30
concat.cu
musa: fix all warnings, re-enable -DLLAMA_FATAL_WARNINGS=ON in ci and update doc (#12611)
2025-03-30 10:59:38 +02:00
concat.cuh
llama : reorganize source code + improve CMake (#8006)
2024-06-26 18:33:02 +03:00
conv2d-dw.cu
CUDA: add conv_2d_dw (#14265)
2025-06-20 09:50:24 +08:00
conv2d-dw.cuh
CUDA: add conv_2d_dw (#14265)
2025-06-20 09:50:24 +08:00
conv2d-transpose.cu
CUDA: add conv_2d_transpose (#14287)
2025-06-20 22:48:24 +08:00
conv2d-transpose.cuh
CUDA: add conv_2d_transpose (#14287)
2025-06-20 22:48:24 +08:00
conv2d.cu
CUDA: fix build error from ambiguous __half conversions in conv2d (#15690)
2025-09-01 06:55:06 +05:30
conv2d.cuh
CUDA: add conv2d (#15635)
2025-08-28 20:33:03 +02:00
conv-transpose-1d.cu
musa: add GGML_UNUSED_VARS (#15446)
2025-08-21 11:06:05 +08:00
conv-transpose-1d.cuh
feat: cuda implementation for ggml_conv_transpose_1d (ggml/854)
2024-07-08 12:23:00 +03:00
convert.cu
CUDA: loop over ne2*ne3 in case it overflows (#19538)
2026-02-13 17:01:40 +05:30
convert.cuh
HIP: RDNA4 tensor core support for MMF (#17077)
2025-11-22 00:03:24 +01:00
count-equal.cu
ggml: fix zero division in ‘dne’ calculation in CUDA COUNT_EQUAL operator when ‘ne’ is small (#10213)
2024-11-09 08:35:46 +01:00
count-equal.cuh
ggml/ex: calculate accuracy in graph, adapt MNIST (ggml/980)
2024-10-03 21:17:26 +03:00
cp-async.cuh
CUDA: FA support for Deepseek (Ampere or newer) (#13306)
2025-05-09 13:34:58 +02:00
cpy-utils.cuh
cuda : support non-contiguous i32 to i32 copy (#17326)
2025-11-23 11:13:34 +01:00
cpy.cu
cuda : fix copy of large tensors (ggml_nbytes <= INT_MAX assertion) (#18433)
2026-01-02 00:24:20 +01:00
cpy.cuh
cuda : remove legacy copy-op pointer indirection code (#16485)
2025-10-14 11:53:49 +02:00
cross-entropy-loss.cu
CUDA: add dynamic shared mem to softmax, refactor general usage (#14497)
2025-07-03 07:45:11 +08:00
cross-entropy-loss.cuh
ggml/examples: add backend support for numerical optimization (ggml/949)
2024-09-20 21:15:05 +03:00
cumsum.cu
sampling : add support for backend sampling (#17004)
2026-01-04 22:22:16 +02:00
cumsum.cuh
Add support for CUMSUM and TRI for CUDA. (#17584)
2025-12-04 22:19:51 +01:00
dequantize.cuh
CUDA: replace GGML_CUDA_F16 with CUDA arch checks (#15433)
2025-08-20 16:58:49 +02:00
diag.cu
Add DIAG for CUDA (#17873)
2025-12-09 20:28:57 +01:00
diag.cuh
Add DIAG for CUDA (#17873)
2025-12-09 20:28:57 +01:00
diagmask.cu
llama : reorganize source code + improve CMake (#8006)
2024-06-26 18:33:02 +03:00
diagmask.cuh
llama : reorganize source code + improve CMake (#8006)
2024-06-26 18:33:02 +03:00
fattn-common.cuh
cuda : fix "V is K view" check for non-unified KV cache (#19145)
2026-01-28 09:15:27 +02:00
fattn-mma-f16.cuh
CUDA: add CDNA3 MFMA support for flash attention MMA kernel (#19806)
2026-02-27 19:37:26 +01:00
fattn-tile.cu
ggml: CUDA: add head size 72 for flash-attn (#16962)
2025-11-03 14:29:11 +01:00
fattn-tile.cuh
CUDA: fix kernel selection logic for tile FA (#19686)
2026-02-19 12:42:58 +01:00
fattn-vec.cuh
CUDA: fix allignment on register spill for FA (#18815)
2026-01-15 15:14:50 +01:00
fattn-wmma-f16.cu
Adjust workaround for ROCWMMA_FATTN/GFX9 to only newer ROCm veresions (#19591)
2026-02-16 14:46:08 +01:00
fattn-wmma-f16.cuh
CUDA: generalized (mma) FA, add Volta support (#17505)
2025-12-03 16:57:05 +01:00
fattn.cu
CUDA: add CDNA3 MFMA support for flash attention MMA kernel (#19806)
2026-02-27 19:37:26 +01:00
fattn.cuh
CUDA: refactor FA support/selection code (#15454)
2025-08-20 23:14:14 +02:00
fill.cu
ggml : allow fill node alloc inplace (#17870)
2025-12-09 12:23:47 +01:00
fill.cuh
cuda : add FILL op support (#17851)
2025-12-08 21:10:12 +08:00
getrows.cu
CUDA: fix GET_ROWS for large tensors (#15882)
2025-09-09 08:11:01 +02:00
getrows.cuh
CUDA: batched+noncont MMQ, refactor bs>1 MoE code (#13199)
2025-04-30 23:12:59 +02:00
ggml-cuda.cu
Improve CUDA graph capture (#19754)
2026-02-21 15:09:36 +05:30
gla.cu
llama: add support for QRWKV6 model architecture (#11001)
2025-01-10 09:58:08 +08:00
gla.cuh
llama: add support for QRWKV6 model architecture (#11001)
2025-01-10 09:58:08 +08:00
im2col.cu
CUDA: fix im2col_3d to respect non-contiguous inputs (views) (#15956)
2025-09-16 00:28:31 +02:00
im2col.cuh
ggml: add ops for WAN video model (cuda && cpu) (#15669)
2025-09-04 10:38:49 +02:00
mean.cu
ggml-cuda: enable cuda-graphs for n-cpu-moe (#18934)
2026-01-24 14:25:20 +08:00
mean.cuh
CUDA: add mean operation (#14313)
2025-06-22 12:39:54 +08:00
mma.cuh
CUDA: add CDNA3 MFMA support for flash attention MMA kernel (#19806)
2026-02-27 19:37:26 +01:00
mmf.cu
HIP: add mmf for CDNA (#18896)
2026-01-29 11:10:53 +01:00
mmf.cuh
HIP: add mmf for CDNA (#18896)
2026-01-29 11:10:53 +01:00
mmid.cu
CUDA: add fp kernel for larger batch size MoE (#16512)
2025-10-14 13:15:15 +02:00
mmid.cuh
CUDA: add fp kernel for larger batch size MoE (#16512)
2025-10-14 13:15:15 +02:00
mmq.cu
HIP: adjust RDNA3.5 MMQ kernel selction logic (#18666)
2026-01-10 17:19:01 +01:00
mmq.cuh
cuda: optimize iq2xxs/iq2xs/iq3xxs dequantization (#19624)
2026-02-15 22:38:42 +05:30
mmvf.cu
CUDA: use mmvq for mul-mat-id for small batch sizes (#18958)
2026-02-03 23:31:23 +08:00
mmvf.cuh
CUDA: use mmvq for mul-mat-id for small batch sizes (#18958)
2026-02-03 23:31:23 +08:00
mmvq.cu
CUDA: use mmvq for mul-mat-id for small batch sizes (#18958)
2026-02-03 23:31:23 +08:00
mmvq.cuh
cuda : enable CUDA graphs for MMID 1 <= BS <= 4 (#19645)
2026-02-17 12:31:49 +02:00
norm.cu
CUDA: Factor out and re-use block_reduce function (#18785)
2026-01-15 10:44:54 +08:00
norm.cuh
CUDA: fuse adds, fuse add with rms norm (#15631)
2025-08-29 11:35:58 +08:00
opt-step-adamw.cu
ggml: new optimization interface (ggml/988)
2024-11-17 08:30:29 +02:00
opt-step-adamw.cuh
ggml/examples: add backend support for numerical optimization (ggml/949)
2024-09-20 21:15:05 +03:00
opt-step-sgd.cu
finetune: SGD optimizer, more CLI args (#13873)
2025-08-14 12:03:57 +02:00
opt-step-sgd.cuh
finetune: SGD optimizer, more CLI args (#13873)
2025-08-14 12:03:57 +02:00
out-prod.cu
CPU/CUDA: fix (GQA) mul mat back, add CUDA support (#11380)
2025-01-24 12:38:31 +01:00
out-prod.cuh
ggml/examples: add backend support for numerical optimization (ggml/949)
2024-09-20 21:15:05 +03:00
pad_reflect_1d.cu
musa: fix build warnings (#15611)
2025-09-26 02:56:10 +02:00
pad_reflect_1d.cuh
cuda : add Pad Reflect 1D support (#14659)
2025-08-22 13:06:29 +02:00
pad.cu
cuda : extend GGML_OP_PAD to work with non-cont src0 (#19429)
2026-02-10 08:07:16 +02:00
pad.cuh
llama : reorganize source code + improve CMake (#8006)
2024-06-26 18:33:02 +03:00
pool2d.cu
llama : reorganize source code + improve CMake (#8006)
2024-06-26 18:33:02 +03:00
pool2d.cuh
llama : reorganize source code + improve CMake (#8006)
2024-06-26 18:33:02 +03:00
quantize.cu
CUDA: experimental native mxfp4 support for blackwell (#17906)
2025-12-24 22:28:26 +08:00
quantize.cuh
CUDA: experimental native mxfp4 support for blackwell (#17906)
2025-12-24 22:28:26 +08:00
reduce_rows.cuh
CUDA: Factor out and re-use block_reduce function (#18785)
2026-01-15 10:44:54 +08:00
roll.cu
CUDA: add roll (#14919)
2025-07-29 14:45:18 +08:00
roll.cuh
CUDA: add roll (#14919)
2025-07-29 14:45:18 +08:00
rope.cu
CUDA: Fix non-contig rope (#19338)
2026-02-08 15:12:51 +02:00
rope.cuh
CUDA: fuse rope + set_rows (#16884)
2025-11-13 08:50:01 +08:00
scale.cu
ggml: add ops for WAN video model (cuda && cpu) (#15669)
2025-09-04 10:38:49 +02:00
scale.cuh
llama : reorganize source code + improve CMake (#8006)
2024-06-26 18:33:02 +03:00
set-rows.cu
CUDA: use fastdiv in set-rows (#16834)
2025-10-29 21:11:53 +08:00
set-rows.cuh
CUDA: add set rows for f32 and f16 (#14551)
2025-07-12 16:31:38 +03:00
set.cu
cuda: add SET operation support (#16804)
2025-10-28 20:10:28 +01:00
set.cuh
cuda: add SET operation support (#16804)
2025-10-28 20:10:28 +01:00
softcap.cu
cuda : add softcap fusion (#14907)
2025-07-29 14:22:03 +02:00
softcap.cuh
cuda : add softcap fusion (#14907)
2025-07-29 14:22:03 +02:00
softmax.cu
CUDA: Factor out and re-use block_reduce function (#18785)
2026-01-15 10:44:54 +08:00
softmax.cuh
CUDA: backwards pass for misc. ops, add tests (#11257)
2025-01-16 16:43:38 +01:00
solve_tri.cu
SOLVE_TRI extension to more dimensions (#17793)
2025-12-11 17:20:43 +01:00
solve_tri.cuh
SOLVE_TRI CUDA kernel for small matrices (#17457)
2025-11-28 12:15:32 +08:00
ssm-conv.cu
model : add ASR support for LFM2-Audio-1.5B (conformer) (#18106)
2025-12-19 00:18:01 +01:00
ssm-conv.cuh
ggml : faster ssm scan (#10558)
2025-03-31 18:05:13 +02:00
ssm-scan.cu
ggml : optimize cuda ssm_scan using warp-level reduction (#18505)
2026-01-07 02:24:34 +08:00
ssm-scan.cuh
ggml : faster ssm scan (#10558)
2025-03-31 18:05:13 +02:00
sum.cu
CUDA: Optimize reduce_rows_f32 kernel, leading up to 25x perf improvement on kernel-level and 10% perf increase for Gemma3n (#15132)
2025-08-13 10:04:46 +02:00
sum.cuh
tests: add gradient tests for all backends (ggml/932)
2024-09-08 11:05:55 +03:00
sumrows.cu
CUDA: Optimize reduce_rows_f32 kernel, leading up to 25x perf improvement on kernel-level and 10% perf increase for Gemma3n (#15132)
2025-08-13 10:04:46 +02:00
sumrows.cuh
CUDA: add mean operation (#14313)
2025-06-22 12:39:54 +08:00
top-k.cu
CUDA: Replace init_offsets kernel with iterators in cub-based argsort (#18930)
2026-01-20 20:11:01 +08:00
top-k.cuh
sampling : add support for backend sampling (#17004)
2026-01-04 22:22:16 +02:00
topk-moe.cu
CUDA: refactor topk-moe to enable more models (GLM 4.7, Nemotron etc.) (#19126)
2026-01-29 10:31:28 +08:00
topk-moe.cuh
CUDA: refactor topk-moe to enable more models (GLM 4.7, Nemotron etc.) (#19126)
2026-01-29 10:31:28 +08:00
tri.cu
Add support for CUMSUM and TRI for CUDA. (#17584)
2025-12-04 22:19:51 +01:00
tri.cuh
Add support for CUMSUM and TRI for CUDA. (#17584)
2025-12-04 22:19:51 +01:00
tsembd.cu
ggml : fix padding in timestep embedding kernels (#15932)
2025-09-16 15:25:57 +02:00
tsembd.cuh
llama : reorganize source code + improve CMake (#8006)
2024-06-26 18:33:02 +03:00
unary.cu
ggml : add ops SOFTPLUS, EXPM1, TRI, SOLVE_TRI, CUMSUM (#17063)
2025-11-13 20:54:47 +02:00
unary.cuh
ggml : add ops SOFTPLUS, EXPM1, TRI, SOLVE_TRI, CUMSUM (#17063)
2025-11-13 20:54:47 +02:00
upscale.cu
model: LFM2-VL fixes (#17577)
2025-11-30 21:57:31 +01:00
upscale.cuh
llama : reorganize source code + improve CMake (#8006)
2024-06-26 18:33:02 +03:00
vecdotq.cuh
cuda: optimize iq2xxs/iq2xs/iq3xxs dequantization (#19624)
2026-02-15 22:38:42 +05:30
wkv.cu
llama: Add support for RWKV v7 architecture (#12412)
2025-03-18 07:27:50 +08:00
wkv.cuh
llama: Add support for RWKV v7 architecture (#12412)
2025-03-18 07:27:50 +08:00