diff --git a/tests/v1/attention/test_rocm_aiter_mla_sparse.py b/tests/v1/attention/test_rocm_aiter_mla_sparse.py new file mode 100644 index 0000000000..1d28b57beb --- /dev/null +++ b/tests/v1/attention/test_rocm_aiter_mla_sparse.py @@ -0,0 +1,54 @@ +# SPDX-License-Identifier: Apache-2.0 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project + +from types import SimpleNamespace + +import pytest +import torch + +from vllm._aiter_ops import is_aiter_found_and_supported +from vllm.platforms import current_platform + +if not current_platform.is_rocm() or not is_aiter_found_and_supported(): + pytest.skip( + "ROCm AITER sparse indexer test requires ROCm with AITER.", + allow_module_level=True, + ) + +from vllm._aiter_ops import rocm_aiter_ops +from vllm.v1.attention.ops import rocm_aiter_mla_sparse as sparse_mod + + +def test_rocm_fp8_mqa_logits_disables_clean_logits(monkeypatch): + captured = {} + expected = torch.empty((1, 1)) + + def fake_fp8_mqa_logits(*args, clean_logits): + captured["clean_logits"] = clean_logits + return expected + + monkeypatch.setattr(sparse_mod, "_ON_GFX942", False) + monkeypatch.setattr(rocm_aiter_ops, "is_enabled", lambda: True) + monkeypatch.setattr( + sparse_mod, + "mqa_logits_module", + lambda: SimpleNamespace(fp8_mqa_logits=fake_fp8_mqa_logits), + ) + + q = torch.empty((1, 1, 1)) + k_fp8 = torch.empty((1, 1)) + scale = torch.empty((1,)) + weights = torch.empty((1, 1)) + cu_seqlen_ks = torch.zeros((1,), dtype=torch.int32) + cu_seqlen_ke = torch.ones((1,), dtype=torch.int32) + + result = sparse_mod.rocm_fp8_mqa_logits( + q, + (k_fp8, scale), + weights, + cu_seqlen_ks, + cu_seqlen_ke, + ) + + assert result is expected + assert captured["clean_logits"] is False diff --git a/vllm/v1/attention/ops/rocm_aiter_mla_sparse.py b/vllm/v1/attention/ops/rocm_aiter_mla_sparse.py index 6a93d21511..78ce85f6ae 100644 --- a/vllm/v1/attention/ops/rocm_aiter_mla_sparse.py +++ b/vllm/v1/attention/ops/rocm_aiter_mla_sparse.py @@ -933,7 +933,9 @@ def rocm_fp8_mqa_logits( if aiter_mqa_logits_module is not None: fp8_mqa_logits = aiter_mqa_logits_module.fp8_mqa_logits - return fp8_mqa_logits(q, k_fp8, scale, weights, cu_seqlen_ks, cu_seqlen_ke) + return fp8_mqa_logits( + q, k_fp8, scale, weights, cu_seqlen_ks, cu_seqlen_ke, clean_logits=False + ) else: return fp8_mqa_logits_torch(q, kv, weights, cu_seqlen_ks, cu_seqlen_ke)