frozenleaves/vllm
1
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2025-10-20 14:53:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

[BugFix] Illegal Memory Access in the blockwise cutlass fp8 GEMMs (#14396)

Browse Source
This commit is contained in:
Lucas Wilkinson
2025-03-07 00:56:06 -05:00
committed by GitHub
parent e1744502c2
commit e5e03c2c1b
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
csrc/cutlass_extensions/gemm/collective/sm90_mma_tma_gmma_ss_warpspecialized_fp8_blockwise_scaling.hpp
View File

@ -402,7 +402,7 @@ struct CollectiveMma<
// TODO: test `scale_copy_a` with `ScaleMsPerTile` < 128
TiledCopy scale_copy_a = make_tiled_copy(SmemBlockScalingCopyAtomA{},
Layout<Shape<_32, _1>>{}, Layout<Shape<_4, _1>>{}); // (1,1,1)
Layout<Shape<_32>>{}, Layout<Shape<_1>>{}); // (1,1,1)
TiledCopy scale_copy_b = make_tiled_copy(SmemBlockScalingCopyAtomB{},
Layout<Shape<_1>>{}, Layout<Shape<_1>>{}); // (1,1,1)
ThrCopy thr_scale_copy_a = scale_copy_a.get_slice(threadIdx.x);