[PERF]MiniMax-M2 gate kernel (#38445)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Signed-off-by: qianlihuang <91178480+qianlihuang@users.noreply.github.com>
Co-authored-by: Yiliu Dong <91178480+qianlihuang@users.noreply.github.com>
This commit is contained in:
Jee Jee Li
2026-05-29 18:28:34 -07:00
committed by GitHub
co-authored by Yiliu Dong
parent 187457a952
commit 559d6710bf
10 changed files with 716 additions and 23 deletions
+24 -10
View File
@@ -683,6 +683,22 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
"in CUDA target architectures.")
endif()
# FP32 router GEMM (H=3072, E=256, M<=32). Requires SM90+ and CUDA >= 12.0.
cuda_archs_sm90plus(FP32_ROUTER_GEMM_ARCHS "${CUDA_ARCHS}")
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND FP32_ROUTER_GEMM_ARCHS)
set(SRCS
"csrc/libtorch_stable/fp32_router_gemm_entry.cu"
"csrc/libtorch_stable/fp32_router_gemm.cu")
set_gencode_flags_for_srcs(
SRCS "${SRCS}"
CUDA_ARCHS "${FP32_ROUTER_GEMM_ARCHS}")
list(APPEND VLLM_STABLE_EXT_SRC "${SRCS}")
message(STATUS "Building fp32_router_gemm for archs: ${FP32_ROUTER_GEMM_ARCHS}")
else()
message(STATUS "Not building fp32_router_gemm as no compatible archs found "
"(requires SM90+ and CUDA >= 12.0).")
endif()
# Only build AllSpark kernels if we are building for at least some compatible archs.
cuda_archs_loose_intersection(ALLSPARK_ARCHS "8.0;8.6;8.7;8.9" "${CUDA_ARCHS}")
if (ALLSPARK_ARCHS)
@@ -1240,24 +1256,22 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
" in CUDA target architectures")
endif()
# DeepSeek V3 router GEMM kernel - requires SM90+
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
cuda_archs_loose_intersection(DSV3_ROUTER_GEMM_ARCHS "9.0a;10.0f;11.0f" "${CUDA_ARCHS}")
else()
cuda_archs_loose_intersection(DSV3_ROUTER_GEMM_ARCHS "9.0a;10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
endif()
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND DSV3_ROUTER_GEMM_ARCHS)
# DeepSeek V3 router GEMM kernel requires SM90+ and CUDA >= 12.0.
# (fp32_router_gemm has been migrated to _C_stable_libtorch above.)
cuda_archs_sm90plus(SM90PLUS_ROUTER_GEMM_ARCHS "${CUDA_ARCHS}")
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SM90PLUS_ROUTER_GEMM_ARCHS)
set(DSV3_ROUTER_GEMM_SRC
"csrc/moe/dsv3_router_gemm_entry.cu"
"csrc/moe/dsv3_router_gemm_float_out.cu"
"csrc/moe/dsv3_router_gemm_bf16_out.cu")
set_gencode_flags_for_srcs(
SRCS "${DSV3_ROUTER_GEMM_SRC}"
CUDA_ARCHS "${DSV3_ROUTER_GEMM_ARCHS}")
CUDA_ARCHS "${SM90PLUS_ROUTER_GEMM_ARCHS}")
list(APPEND VLLM_MOE_EXT_SRC "${DSV3_ROUTER_GEMM_SRC}")
message(STATUS "Building DSV3 router GEMM kernel for archs: ${DSV3_ROUTER_GEMM_ARCHS}")
message(STATUS "Building DSV3 router GEMM kernels for archs: ${SM90PLUS_ROUTER_GEMM_ARCHS}")
else()
message(STATUS "Not building DSV3 router GEMM kernel as no compatible archs found"
message(STATUS "Not building DSV3 router GEMM kernels as no compatible archs found"
" (requires SM90+ and CUDA >= 12.0)")
endif()
endif()