Skip to content

Commit c2f3e04

Browse files
author
da.huo
committed
restore
1 parent 59b2264 commit c2f3e04

1 file changed

Lines changed: 2 additions & 5 deletions

File tree

src/turbomind/kernels/gemm/CMakeLists.txt

Lines changed: 2 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -85,11 +85,8 @@ set_property(TARGET gemm2 PROPERTY POSITION_INDEPENDENT_CODE ON)
8585
set_property(TARGET gemm2 PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS ON)
8686

8787
if(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

0 commit comments

Comments
 (0)