File tree Expand file tree Collapse file tree
src/turbomind/kernels/gemm Expand file tree Collapse file tree Original file line number Diff line number Diff line change @@ -85,11 +85,8 @@ set_property(TARGET gemm2 PROPERTY POSITION_INDEPENDENT_CODE ON)
8585set_property (TARGET gemm2 PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS ON )
8686
8787if (GEMM2_ARCH_90_ENABLED)
88- # SM90 kernels compile only for sm_90 (CUTLASS wgmma instructions).
89- # tma.cu is duplicated here (also in gemm2) so make_2d_tma_desc lives in the
90- # same archive as its only callers (kernel_impl_sm90.h), avoiding the undefined
91- # symbol from single-pass static-link ordering between two archives.
92- add_library (gemm2_sm90 STATIC ${GEMM2_KERNELS_SM90} tma.cu )
88+ # SM90 kernels only compile for 90/90a; avoid building them for sm_100.
89+ add_library (gemm2_sm90 STATIC ${GEMM2_KERNELS_SM90} )
9390 set_target_properties (gemm2_sm90 PROPERTIES
9491 CUDA_ARCHITECTURES "${_sm90_archs} "
9592 POSITION_INDEPENDENT_CODE ON
You can’t perform that action at this time.
0 commit comments