Skip to content

Commit eb36918

Browse files
henrylhtsangfacebook-github-bot
authored andcommitted
re-enable decode tests (#4946)
Summary: X-link: facebookresearch/FBGEMM#1966 Pull Request resolved: #4946 Previously we disabled the test because it would crash. It seemed to be caused by tensorwise scaled_mm, but we couldn't bisect. After buck2 clean, it seems working again. Also asking in https://fb.workplace.com/groups/hm.sup/posts/1902458160316164/, in case it was caused by a driver upgrade. Reviewed By: q10 Differential Revision: D83501117 fbshipit-source-id: 09c97a197f344490204ae7386dc005ce599026a4
1 parent a4c1ecd commit eb36918

File tree

1 file changed

+0
-1
lines changed

1 file changed

+0
-1
lines changed

fbgemm_gpu/experimental/gen_ai/test/attention/blackwell_fmha_test.py

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -431,7 +431,6 @@ def _execute_cutlass_blackwell_attn_varlen(
431431
assert torch.equal(dk, dk_d)
432432
assert torch.equal(dv, dv_d)
433433

434-
@unittest.skipIf(True, "Skip this until we figure out an issue with scaled_mm")
435434
@skip_cuda_lt_sm100
436435
@skip_rocm
437436
@parameterized.expand(

0 commit comments

Comments
 (0)