diff --git a/tests/kernels/attention/test_rocm_triton_attn_dsv4.py b/tests/kernels/attention/test_rocm_triton_attn_dsv4.py index 362ac6b7ce..51b77ce172 100644 --- a/tests/kernels/attention/test_rocm_triton_attn_dsv4.py +++ b/tests/kernels/attention/test_rocm_triton_attn_dsv4.py @@ -70,7 +70,7 @@ def enable_aiter_mqa(monkeypatch): def test_unpaged_mqa_logits_preserves_intervals_and_scale_layout( column_scales, monkeypatch, enable_aiter_mqa ) -> None: - """The vLLM AITER dispatch preserves ragged masking and per-token scales.""" + """The vLLM AITER dispatch honors ragged intervals and per-token scales.""" from vllm._aiter_ops import rocm_aiter_ops from vllm.v1.attention.ops import rocm_aiter_mla_sparse as mod @@ -108,7 +108,7 @@ def test_unpaged_mqa_logits_preserves_intervals_and_scale_layout( valid = (positions >= starts[:, None]) & (positions < ends[:, None]) assert actual.shape == expected.shape assert actual.dtype == torch.float32 - assert torch.equal(torch.isneginf(actual), ~valid) + assert not torch.isneginf(actual[valid]).any() torch.testing.assert_close(actual[valid], expected[valid], rtol=1e-3, atol=1e-3)