| 12345678910111213141516171819202122232425262728 |
- Fix: deterministically honor TORCH_CUDA_ARCH_LIST for the vLLM GPU arch.
- Deriving CUDA_ARCHS from the torch wheel arch list is unreliable for a DGX Spark
- build: the stock cu130 torch advertises sm_120, and the sm_121a-only FP4
- instruction cvt.e2m1x2.f32 then fails in ptxas. When TORCH_CUDA_ARCH_LIST is set
- (e.g. 12.1a), take the numeric arch (12.1) directly so the CUDA>=13 family
- mapping rebuilds for sm_121a.
- Note: this is one of two adjacent changes; see 0001-cuda13-supported-archs-add-12.1.patch.
- ---
- --- a/CMakeLists.txt
- +++ b/CMakeLists.txt
- @@ -221,6 +221,16 @@
- #
- clear_cuda_arches(CUDA_ARCH_FLAGS)
- extract_unique_cuda_archs_ascending(CUDA_ARCHS "${CUDA_ARCH_FLAGS}")
- + # Deterministically honor TORCH_CUDA_ARCH_LIST (e.g. "12.1a" -> 12.1)
- + # instead of deriving from the torch wheel arch list. Required so a
- + # DGX Spark (GB10/sm_121a) build targets the 12.1 family; otherwise the
- + # sm_120 FP4 cvt.e2m1x2.f32 path fails in ptxas.
- + if(DEFINED ENV{TORCH_CUDA_ARCH_LIST} AND NOT "$ENV{TORCH_CUDA_ARCH_LIST}" STREQUAL "")
- + string(REGEX MATCH "[0-9]+\.[0-9]+" _forced_arch "$ENV{TORCH_CUDA_ARCH_LIST}")
- + if(NOT _forced_arch STREQUAL "")
- + set(CUDA_ARCHS "${_forced_arch}")
- + endif()
- + endif()
- message(STATUS "CUDA target architectures: ${CUDA_ARCHS}")
- # Filter the target architectures by the supported supported archs
- # since for some files we will build for all CUDA_ARCHS.
|
