ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

CUDA: Improve flash decoding kernel GPU occupancy for BS=1 case (#12183)

Browse source 
- Find out active blocks per SM using cudaOccupancyMaxActiveBlocksPerMultiprocessor API. Use this value to determine the optimal parallel_blocks value.
- Prefer vector flash attention kernels over MMA kernel for BS=1

Fixes Issue: #12182
---------

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
This commit is contained in:
Gaurav Garg 2025-03-20 01:22:06 +05:30 • committed by GitHub
parent a9b59288e2
commit 517b5ddbf0
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
12 changed files with 214 additions and 266 deletions
Download patch file Download diff file Expand all files Collapse all files

3
ggml/src/ggml-cuda/fattn-mma-f16.cuh
View file

@ -970,7 +970,8 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
fattn_kernel = flash_attn_ext_f16<D, ncols1, ncols2, nwarps, KQ_per_iter, ntiles, use_logit_softcap>;
}
launch_fattn<D, ncols1, ncols2, 0, KQ_per_iter>(ctx, dst, fattn_kernel, nwarps, nbytes_shared_total, true, true);
launch_fattn<D, ncols1, ncols2, KQ_per_iter>
(ctx, dst, fattn_kernel, nwarps, nbytes_shared_total, FATTN_KQ_STRIDE, true, true, true);
}