[Kernel] CUTLASS grouped gemm fp8 MoE kernel (#13972)

Signed-off-by: ElizaWszola <eliza@neuralmagic.com>
Signed-off-by: ElizaWszola <ewszola@redhat.com>
Co-authored-by: Lucas Wilkinson <wilkinson.lucas@gmail.com>
This commit is contained in:
ElizaWszola
2025-03-27 00:54:44 +00:00
committed by GitHub
co-authored by Lucas Wilkinson
parent 7a6d45bc8a
commit 9239bf718e
22 changed files with 2317 additions and 15 deletions
+27
View File
@@ -461,6 +461,33 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
set(FP4_ARCHS)
endif()
#
# CUTLASS MoE kernels
# The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and only works
# on Hopper). get_cutlass_moe_mm_data should only be compiled if it's possible
# to compile MoE kernels that use its output.
cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu"
"csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
set_gencode_flags_for_srcs(
SRCS "${SRCS}"
CUDA_ARCHS "${SCALED_MM_ARCHS}")
list(APPEND VLLM_EXT_SRC "${SRCS}")
list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM90=1")
message(STATUS "Building grouped_mm_c3x for archs: ${SCALED_MM_ARCHS}")
else()
if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
message(STATUS "Not building grouped_mm_c3x kernels as CUDA Compiler version is "
"not >= 12.3, we recommend upgrading to CUDA 12.3 or later "
"if you intend on running FP8 quantized MoE models on Hopper.")
else()
message(STATUS "Not building grouped_mm_c3x as no compatible archs found "
"in CUDA target architectures")
endif()
endif()
#
# Machete kernels