add cutlass support for blackwell fp8 gemm (#13798)
This commit is contained in:
+24
-8
@@ -31,7 +31,7 @@ set(ignoreMe "${VLLM_PYTHON_PATH}")
|
||||
set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12")
|
||||
|
||||
# Supported NVIDIA architectures.
|
||||
set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0")
|
||||
set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0")
|
||||
|
||||
# Supported AMD GPU architectures.
|
||||
set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101")
|
||||
@@ -297,7 +297,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
|
||||
# Only build Marlin kernels if we are building for at least some compatible archs.
|
||||
# Keep building Marlin for 9.0 as there are some group sizes and shapes that
|
||||
# are not supported by Machete yet.
|
||||
cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.7;8.9;9.0" "${CUDA_ARCHS}")
|
||||
cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
|
||||
if (MARLIN_ARCHS)
|
||||
set(MARLIN_SRCS
|
||||
"csrc/quantization/fp8/fp8_marlin.cu"
|
||||
@@ -335,7 +335,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
|
||||
|
||||
# The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
|
||||
# CUDA 12.0 or later (and only work on Hopper, 9.0a for now).
|
||||
cuda_archs_loose_intersection(SCALED_MM_3X_ARCHS "9.0a" "${CUDA_ARCHS}")
|
||||
cuda_archs_loose_intersection(SCALED_MM_3X_ARCHS "9.0a;10.0a;10.1a;12.0a" "${CUDA_ARCHS}")
|
||||
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_3X_ARCHS)
|
||||
set(SRCS
|
||||
"csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu"
|
||||
@@ -369,7 +369,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
|
||||
# For the cutlass_scaled_mm kernels we want to build the c2x (CUTLASS 2.x)
|
||||
# kernels for the remaining archs that are not already built for 3x.
|
||||
cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS
|
||||
"7.5;8.0;8.6;8.7;8.9;9.0" "${CUDA_ARCHS}")
|
||||
"7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
|
||||
# subtract out the archs that are already built for 3x
|
||||
list(REMOVE_ITEM SCALED_MM_2X_ARCHS ${SCALED_MM_3X_ARCHS})
|
||||
if (SCALED_MM_2X_ARCHS)
|
||||
@@ -394,7 +394,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
|
||||
# 2:4 Sparse Kernels
|
||||
|
||||
# The 2:4 sparse kernels cutlass_scaled_sparse_mm and cutlass_compressor
|
||||
# require CUDA 12.2 or later (and only work on Hopper, 9.0a for now).
|
||||
# require CUDA 12.2 or later (and only work on Hopper and Blackwell).
|
||||
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_3X_ARCHS)
|
||||
set(SRCS "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
|
||||
set_gencode_flags_for_srcs(
|
||||
@@ -419,8 +419,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
|
||||
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND FP4_ARCHS)
|
||||
set(SRCS
|
||||
"csrc/quantization/fp4/nvfp4_quant_kernels.cu"
|
||||
"csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu"
|
||||
)
|
||||
"csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu")
|
||||
set_gencode_flags_for_srcs(
|
||||
SRCS "${SRCS}"
|
||||
CUDA_ARCHS "${FP4_ARCHS}")
|
||||
@@ -433,6 +432,22 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
|
||||
set(FP4_ARCHS)
|
||||
endif()
|
||||
|
||||
# FP8 Blackwell Archs
|
||||
cuda_archs_loose_intersection(BLACKWELL_ARCHS "10.0;10.1;12.0" "${CUDA_ARCHS}")
|
||||
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND BLACKWELL_ARCHS)
|
||||
set(SRCS
|
||||
"csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
|
||||
)
|
||||
set_gencode_flags_for_srcs(
|
||||
SRCS "${SRCS}"
|
||||
CUDA_ARCHS "${BLACKWELL_ARCHS}")
|
||||
list(APPEND VLLM_EXT_SRC "${SRCS}")
|
||||
message(STATUS "Building FP8 for archs: ${BLACKWELL_ARCHS}")
|
||||
else()
|
||||
# clear BLACKWELL_ARCHS
|
||||
set(BLACKWELL_ARCHS)
|
||||
endif()
|
||||
|
||||
#
|
||||
# Machete kernels
|
||||
|
||||
@@ -514,6 +529,7 @@ define_gpu_extension_target(
|
||||
COMPILE_FLAGS ${VLLM_GPU_FLAGS}
|
||||
ARCHITECTURES ${VLLM_GPU_ARCHES}
|
||||
INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
|
||||
INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
|
||||
USE_SABI 3
|
||||
WITH_SOABI)
|
||||
|
||||
@@ -537,7 +553,7 @@ set_gencode_flags_for_srcs(
|
||||
CUDA_ARCHS "${CUDA_ARCHS}")
|
||||
|
||||
if(VLLM_GPU_LANG STREQUAL "CUDA")
|
||||
cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.7;8.9;9.0" "${CUDA_ARCHS}")
|
||||
cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
|
||||
if (MARLIN_MOE_ARCHS)
|
||||
set(MARLIN_MOE_SRC
|
||||
"csrc/moe/marlin_kernels/marlin_moe_kernel.h"
|
||||
|
||||
Reference in New Issue
Block a user