frozenleaves/vllm-dev
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

SM100 Cutlass MLA decode with unrestricted num_heads (< 128) for DeepSeek TP (#20769)

Browse Source 
Signed-off-by: Alexander Matveev <amatveev@redhat.com>
This commit is contained in:
Alexander Matveev
2025-07-14 21:06:38 -04:00
committed by GitHub
parent 61e20828da
commit 8cdc371217
12 changed files with 3283 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
CMakeLists.txt
View File

@ -553,7 +553,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
cuda_archs_loose_intersection(MLA_ARCHS "10.0a" "${CUDA_ARCHS}")
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND MLA_ARCHS)
set(SRCS
"csrc/attention/mla/cutlass_mla_kernels.cu")
"csrc/attention/mla/cutlass_mla_kernels.cu"
"csrc/attention/mla/sm100_cutlass_mla_kernel.cu")
set_gencode_flags_for_srcs(
SRCS "${SRCS}"
CUDA_ARCHS "${MLA_ARCHS}")