From 5a6c7b7ab569f49491b5428a7983be5b17b85378 Mon Sep 17 00:00:00 2001 From: Wentao Ye <44945378+yewentao256@users.noreply.github.com> Date: Thu, 11 Jun 2026 16:22:26 -0400 Subject: [PATCH] [Bug] Fix test flashmla for DSv4 (#45052) Signed-off-by: yewentao256 --- tests/kernels/attention/test_flashmla_sparse.py | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/tests/kernels/attention/test_flashmla_sparse.py b/tests/kernels/attention/test_flashmla_sparse.py index 9e4e7c2ec9..d92dabe9d3 100644 --- a/tests/kernels/attention/test_flashmla_sparse.py +++ b/tests/kernels/attention/test_flashmla_sparse.py @@ -29,8 +29,10 @@ def test_sparse_flashmla_metadata_smoke(): topk=topk, is_fp8_kvcache=True, ) - assert tile_md.dtype == torch.int32 - assert num_splits.dtype == torch.int32 + assert isinstance(tile_md, fm.FlashMLASchedMeta) + assert tile_md.tile_scheduler_metadata is None + assert tile_md.num_splits is None + assert num_splits is None def test_sparse_flashmla_decode_smoke(): @@ -116,7 +118,7 @@ def test_sparse_flashmla_prefill_smoke(): kv = torch.zeros((s_kv, h_kv, d_qk), dtype=torch.bfloat16, device=device) indices = torch.zeros((s_q, h_kv, topk), dtype=torch.int32, device=device) - out, max_logits, lse = fm.flash_mla_sparse_prefill(q, kv, indices, 1.0, d_v) + out, max_logits, lse = fm.flash_mla_sparse_fwd(q, kv, indices, 1.0, d_v) assert out.shape == (s_q, h_q, d_v) assert max_logits.shape == (s_q, h_q) assert lse.shape == (s_q, h_q)