xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-07-20 13:37:20 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Build moe_data for both sm100 and sm90 (#20086)

Browse Source 
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin 2025-06-27 12:50:12 +09:00 committed by GitHub
parent 2d7779f888
commit 44d2e6af63
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 17 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
CMakeLists.txt
View File

@ -513,6 +513,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
CUDA_ARCHS "${FP4_ARCHS}") CUDA_ARCHS "${FP4_ARCHS}")
list(APPEND VLLM_EXT_SRC "${SRCS}") list(APPEND VLLM_EXT_SRC "${SRCS}")
list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4=1") list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4=1")
list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
message(STATUS "Building NVFP4 for archs: ${FP4_ARCHS}") message(STATUS "Building NVFP4 for archs: ${FP4_ARCHS}")
else() else()
message(STATUS "Not building NVFP4 as no compatible archs were found.") message(STATUS "Not building NVFP4 as no compatible archs were found.")
@ -547,8 +548,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
# if it's possible to compile MoE kernels that use its output. # if it's possible to compile MoE kernels that use its output.
cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${CUDA_ARCHS}") 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) if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu" set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu")
"csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
set_gencode_flags_for_srcs( set_gencode_flags_for_srcs(
SRCS "${SRCS}" SRCS "${SRCS}"
CUDA_ARCHS "${SCALED_MM_ARCHS}") CUDA_ARCHS "${SCALED_MM_ARCHS}")
@ -566,6 +566,16 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
endif() endif()
endif() endif()
# moe_data.cu is used by all CUTLASS MoE kernels.
cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a" "${CUDA_ARCHS}")
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
set(SRCS "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
set_gencode_flags_for_srcs(
SRCS "${SRCS}"
CUDA_ARCHS "${CUTLASS_MOE_DATA_ARCHS}")
list(APPEND VLLM_EXT_SRC "${SRCS}")
endif()
# #
# Machete kernels # Machete kernels

9
csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu
View File

@ -241,7 +241,7 @@ void get_cutlass_moe_mm_data(
// mm to run it for. // mm to run it for.
int32_t version_num = get_sm_version_num(); int32_t version_num = get_sm_version_num();
#if (defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90) || \ #if (defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90) || \
(defined ENABLE_SCALED_MM_SM100 && ENABLE_SCALED_MM_SM90) (defined ENABLE_CUTLASS_MOE_SM100 && ENABLE_CUTLASS_MOE_SM100)
get_cutlass_moe_mm_data_caller(topk_ids, expert_offsets, problem_sizes1, get_cutlass_moe_mm_data_caller(topk_ids, expert_offsets, problem_sizes1,
problem_sizes2, input_permutation, problem_sizes2, input_permutation,
output_permutation, num_experts, n, k, output_permutation, num_experts, n, k,
@ -252,7 +252,7 @@ void get_cutlass_moe_mm_data(
false, false,
"No compiled get_cutlass_moe_mm_data: no cutlass_scaled_mm kernel for " "No compiled get_cutlass_moe_mm_data: no cutlass_scaled_mm kernel for "
"CUDA device capability: ", "CUDA device capability: ",
version_num, ". Required capability: 90"); version_num, ". Required capability: 90 or 100");
} }
void get_cutlass_pplx_moe_mm_data(torch::Tensor& expert_offsets, void get_cutlass_pplx_moe_mm_data(torch::Tensor& expert_offsets,
@ -265,7 +265,8 @@ void get_cutlass_pplx_moe_mm_data(torch::Tensor& expert_offsets,
// This function currently gets compiled only if we have a valid cutlass moe // This function currently gets compiled only if we have a valid cutlass moe
// mm to run it for. // mm to run it for.
int32_t version_num = get_sm_version_num(); int32_t version_num = get_sm_version_num();
#if defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90 #if (defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90) || \
(defined ENABLE_CUTLASS_MOE_SM100 && ENABLE_CUTLASS_MOE_SM100)
get_cutlass_pplx_moe_mm_data_caller(expert_offsets, problem_sizes1, get_cutlass_pplx_moe_mm_data_caller(expert_offsets, problem_sizes1,
problem_sizes2, expert_num_tokens, problem_sizes2, expert_num_tokens,
num_local_experts, padded_m, n, k); num_local_experts, padded_m, n, k);
@ -275,7 +276,7 @@ void get_cutlass_pplx_moe_mm_data(torch::Tensor& expert_offsets,
false, false,
"No compiled get_cutlass_pplx_moe_mm_data: no cutlass_scaled_mm kernel " "No compiled get_cutlass_pplx_moe_mm_data: no cutlass_scaled_mm kernel "
"for CUDA device capability: ", "for CUDA device capability: ",
version_num, ". Required capability: 90"); version_num, ". Required capability: 90 or 100");
} }
void cutlass_scaled_mm_azp(torch::Tensor& c, torch::Tensor const& a, void cutlass_scaled_mm_azp(torch::Tensor& c, torch::Tensor const& a,