A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 10, 17:45 UTC
[BugFix] Fix architecture flags to prevent issues on SM103 (#31150)
Signed-off-by: LopezCastroRoberto <robertol.c510@gmail.com>
Roberto L. Castro committed
Jan 5, 2026 at 21:11 UTC
fdcc5176beb70e801d9baae14f992819b1ffcffb
2 files changed
+13
-5
cmake/external_projects/qutlass.cmake
+8
-3
@@ -31,10 +31,15 @@ if(NOT qutlass_SOURCE_DIR)
31
endif()
32
message(STATUS "[QUTLASS] QuTLASS is available at ${qutlass_SOURCE_DIR}")
33
34
-cuda_archs_loose_intersection(QUTLASS_ARCHS "12.0a;10.0a" "${CUDA_ARCHS}")
35
-if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND QUTLASS_ARCHS)
34
+if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
35
+ cuda_archs_loose_intersection(QUTLASS_ARCHS "12.0a;10.0f" "${CUDA_ARCHS}")
36
+else()
37
+ cuda_archs_loose_intersection(QUTLASS_ARCHS "12.0a;10.0a;10.3a" "${CUDA_ARCHS}")
38
+endif()
39
+
40
+if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND QUTLASS_ARCHS)
41
37
- if(QUTLASS_ARCHS MATCHES "10\\.0a")
42
+ if(QUTLASS_ARCHS MATCHES "10\\.(0a|3a|0f)")
43
set(QUTLASS_TARGET_CC 100)
44
elseif(QUTLASS_ARCHS MATCHES "12\\.0a")
45
set(QUTLASS_TARGET_CC 120)
tools/flashinfer-build.sh
+5
-2
@@ -32,9 +32,12 @@ if [[ "${CUDA_VERSION}" == 11.* ]]; then
32
FI_TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9"
33
elif [[ "${CUDA_VERSION}" == 12.[0-7]* ]]; then
34
FI_TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9 9.0a"
35
+elif [[ "${CUDA_VERSION}" == 12.[8-9]* ]]; then
36
+ # CUDA 12.8–12.9
37
+ FI_TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9 9.0a 10.0a 10.3a 12.0"
38
else
36
- # CUDA 12.8+ supports 10.0a and 12.0
37
- FI_TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9 9.0a 10.0a 12.0"
39
+ # CUDA 13.0+
40
+ FI_TORCH_CUDA_ARCH_LIST="7.5 8.0 8.9 9.0a 10.0f 12.0"
41
fi
42
43
echo "🏗️ Building FlashInfer AOT for arches: ${FI_TORCH_CUDA_ARCH_LIST}"