[Build] Fix CUDA arch detection producing kernel-less builds on SM121 (#49904)

This commit is contained in:
Ayushman Singh
2026-07-27 22:01:56 -07:00
committed by GitHub
parent d223c900d8
commit 74587939b1
3 changed files with 40 additions and 9 deletions
+9 -4
View File
@@ -219,10 +219,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
# the set of architectures we want to compile for and remove the from the
# CMAKE_CUDA_FLAGS so that they are not applied globally.
#
# `+PTX` in TORCH_CUDA_ARCH_LIST is not preserved here. It is emitted by torch
# as `code=compute_*`, while extract_unique_cuda_archs_ascending() records only
# `arch=compute_*`. If a kernel really needs PTX, add `+PTX` to that kernel's
# component-specific arch list below.
# `+PTX` in TORCH_CUDA_ARCH_LIST is not preserved here. If a kernel really
# needs PTX, add `+PTX` to that kernel's component-specific arch list below.
#
clear_cuda_arches(CUDA_ARCH_FLAGS)
extract_unique_cuda_archs_ascending(CUDA_ARCHS "${CUDA_ARCH_FLAGS}")
@@ -232,6 +230,13 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
cuda_archs_loose_intersection(CUDA_ARCHS
"${CUDA_SUPPORTED_ARCHS}" "${CUDA_ARCHS}")
message(STATUS "CUDA supported target architectures: ${CUDA_ARCHS}")
if(NOT CUDA_ARCHS)
message(FATAL_ERROR
"No supported CUDA architectures; the build would produce a binary "
"with no usable kernels. Detected gencode flags: ${CUDA_ARCH_FLAGS}; "
"supported: ${CUDA_SUPPORTED_ARCHS}. "
"Set TORCH_CUDA_ARCH_LIST for your GPU (e.g. 12.0).")
endif()
else()
#
# For other GPU targets override the GPU architectures detected by cmake/torch