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

Use NVCC --compress-mode to reduce binary size by 30% (#20694)

Browse Source 
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2025-07-10 10:26:48 +09:00
committed by GitHub
parent 805d62ca88
commit 49e8c7ea25
1 changed files with 19 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
CMakeLists.txt
View File

@ -171,6 +171,13 @@ if(NVCC_THREADS AND VLLM_GPU_LANG STREQUAL "CUDA")
list(APPEND VLLM_GPU_FLAGS "--threads=${NVCC_THREADS}")
endif()
#
# Set nvcc fatbin compression.
#
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND VLLM_GPU_LANG STREQUAL "CUDA")
list(APPEND VLLM_GPU_FLAGS "-Xfatbin" "-compress-all" "-compress-mode=size")
endif()
#
# Use FetchContent for C++ dependencies that are compiled as part of vLLM's build process.
@ -392,7 +399,7 @@ 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
cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
set(SRCS
"csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm90.cu"
"csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
@ -408,7 +415,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
message(STATUS "Building scaled_mm_c3x_sm90 for archs: ${SCALED_MM_ARCHS}")
else()
if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
message(STATUS "Not building scaled_mm_c3x_sm90 as CUDA Compiler version is "
"not >= 12.0, we recommend upgrading to CUDA 12.0 or "
"later if you intend on running FP8 quantized models on "
@ -423,7 +430,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
# The cutlass_scaled_mm kernels for Geforce Blackwell SM120 (c3x, i.e. CUTLASS 3.x) require
# CUDA 12.8 or later
cuda_archs_loose_intersection(SCALED_MM_ARCHS "12.0;12.0a" "${CUDA_ARCHS}")
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
set(SRCS
"csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm120.cu"
"csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm120_fp8.cu"
@ -437,7 +444,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
message(STATUS "Building scaled_mm_c3x_sm120 for archs: ${SCALED_MM_ARCHS}")
else()
if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
message(STATUS "Not building scaled_mm_c3x_sm120 as CUDA Compiler version is "
"not >= 12.8, we recommend upgrading to CUDA 12.8 or "
"later if you intend on running FP8 quantized models on "
@ -452,7 +459,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
# The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
# require CUDA 12.8 or later
cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a" "${CUDA_ARCHS}")
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
set(SRCS
"csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
"csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
@ -467,7 +474,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
message(STATUS "Building scaled_mm_c3x_sm100 for archs: ${SCALED_MM_ARCHS}")
else()
if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
message(STATUS "Not building scaled_mm_c3x_sm100 as CUDA Compiler version is "
"not >= 12.8, we recommend upgrading to CUDA 12.8 or "
"later if you intend on running FP8 quantized models on "
@ -510,7 +517,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
# The 2:4 sparse kernels cutlass_scaled_sparse_mm and cutlass_compressor
# require CUDA 12.2 or later (and only work on Hopper).
cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.2 AND SCALED_MM_ARCHS)
set(SRCS "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
set_gencode_flags_for_srcs(
SRCS "${SRCS}"
@ -519,7 +526,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
list(APPEND VLLM_GPU_FLAGS "-DENABLE_SPARSE_SCALED_MM_C3X=1")
message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_ARCHS}")
else()
if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.2 AND SCALED_MM_ARCHS)
message(STATUS "Not building sparse_scaled_mm_c3x kernels as CUDA Compiler version is "
"not >= 12.2, we recommend upgrading to CUDA 12.2 or later "
"if you intend on running FP8 sparse quantized models on Hopper.")
@ -531,7 +538,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
# FP4 Archs and flags
cuda_archs_loose_intersection(FP4_ARCHS "10.0a" "${CUDA_ARCHS}")
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND FP4_ARCHS)
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
set(SRCS
"csrc/quantization/fp4/nvfp4_quant_kernels.cu"
"csrc/quantization/fp4/nvfp4_experts_quant.cu"
@ -552,7 +559,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
# CUTLASS MLA Archs and flags
cuda_archs_loose_intersection(MLA_ARCHS "10.0a" "${CUDA_ARCHS}")
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND MLA_ARCHS)
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND MLA_ARCHS)
set(SRCS
"csrc/attention/mla/cutlass_mla_kernels.cu")
set_gencode_flags_for_srcs(
@ -641,7 +648,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
# The machete kernels only work on hopper and require CUDA 12.0 or later.
# Only build Machete kernels if we are building for something compatible with sm90a
cuda_archs_loose_intersection(MACHETE_ARCHS "9.0a" "${CUDA_ARCHS}")
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND MACHETE_ARCHS)
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND MACHETE_ARCHS)
#
# For the Machete kernels we automatically generate sources for various
# preselected input type pairs and schedules.
@ -693,7 +700,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
message(STATUS "Building Machete kernels for archs: ${MACHETE_ARCHS}")
else()
if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0
if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0
AND MACHETE_ARCHS)
message(STATUS "Not building Machete kernels as CUDA Compiler version is "
"not >= 12.0, we recommend upgrading to CUDA 12.0 or "