Add sm_107 for Rubin (#49387)
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com> Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
co-authored by
OpenAI Codex
parent
41798069f3
commit
8c13ee5735
+12
-7
@@ -114,6 +114,11 @@ find_package(Torch REQUIRED)
|
||||
# Supported NVIDIA architectures.
|
||||
# This check must happen after find_package(Torch) because that's when CMAKE_CUDA_COMPILER_VERSION gets defined
|
||||
if(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
|
||||
CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13.4)
|
||||
# Rubin (10.7) can run SM100 family code, but CUDA 13.4 also supports
|
||||
# targeting it directly.
|
||||
set(CUDA_SUPPORTED_ARCHS "7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.7;11.0;12.0")
|
||||
elseif(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
|
||||
CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13.0)
|
||||
# starting from CUDA 12.9 and Blackwell (10.0), we use family-specific targets (10.0f, 12.0f, etc)
|
||||
# to support the whole generation without specifying all sub-architectures
|
||||
@@ -420,7 +425,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
|
||||
|
||||
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
|
||||
cuda_archs_loose_intersection(COOPERATIVE_TOPK_ARCHS
|
||||
"9.0a;10.0f;10.1f;10.3f;11.0f;12.0f;12.1f" "${CUDA_ARCHS}")
|
||||
"9.0a;10.0f;10.1f;10.3f;10.7f;11.0f;12.0f;12.1f" "${CUDA_ARCHS}")
|
||||
else()
|
||||
cuda_archs_loose_intersection(COOPERATIVE_TOPK_ARCHS
|
||||
"9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
|
||||
@@ -695,7 +700,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
|
||||
|
||||
# DeepSeek V3 fused A GEMM kernel (requires SM 9.0+, Hopper and later)
|
||||
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
|
||||
cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
|
||||
cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0f;10.7f;11.0f;12.0f" "${CUDA_ARCHS}")
|
||||
else()
|
||||
cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
|
||||
endif()
|
||||
@@ -832,7 +837,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
|
||||
# The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
|
||||
# require CUDA 12.8 or later
|
||||
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
|
||||
cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
|
||||
cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;10.7f;11.0f" "${CUDA_ARCHS}")
|
||||
else()
|
||||
cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
|
||||
endif()
|
||||
@@ -916,7 +921,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
|
||||
endif()
|
||||
|
||||
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
|
||||
cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
|
||||
cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;10.7f;11.0f" "${CUDA_ARCHS}")
|
||||
else()
|
||||
cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
|
||||
endif()
|
||||
@@ -941,7 +946,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
|
||||
|
||||
# moe_data.cu is used by all CUTLASS MoE kernels.
|
||||
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
|
||||
cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
|
||||
cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0f;10.7f;11.0f;12.0f" "${CUDA_ARCHS}")
|
||||
else()
|
||||
cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
|
||||
endif()
|
||||
@@ -998,7 +1003,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
|
||||
# SM10x/11x FP4 kernels. MXFP4 experts quantization is currently compiled
|
||||
# only in this block; SM12x has separate NVFP4 matmul/MoE kernels above.
|
||||
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
|
||||
cuda_archs_loose_intersection(FP4_SM100_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
|
||||
cuda_archs_loose_intersection(FP4_SM100_ARCHS "10.0f;10.7f;11.0f" "${CUDA_ARCHS}")
|
||||
else()
|
||||
cuda_archs_loose_intersection(FP4_SM100_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
|
||||
endif()
|
||||
@@ -1064,7 +1069,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
|
||||
# Runtime dispatch is gated in
|
||||
# vllm/v1/attention/backends/mla/cutlass_mla.py.
|
||||
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
|
||||
cuda_archs_loose_intersection(MLA_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
|
||||
cuda_archs_loose_intersection(MLA_ARCHS "10.0f;10.7f;11.0f" "${CUDA_ARCHS}")
|
||||
else()
|
||||
cuda_archs_loose_intersection(MLA_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
|
||||
endif()
|
||||
|
||||
Reference in New Issue
Block a user