From 6c7e679f048dc6123caecc3985150766e455ff22 Mon Sep 17 00:00:00 2001 From: Shanshan Shen <467638484@qq.com> Date: Wed, 29 Jul 2026 02:08:50 +0800 Subject: [PATCH] [ROCm][Bugfix] Sanitize AITER paged-MQA logits before sparse top-k for DeepSeek-V4 (#49714) Signed-off-by: shen-shanshan <467638484@qq.com> --- .../attention/test_rocm_triton_attn_dsv4.py | 64 +++++++++++++++++++ .../v1/attention/ops/rocm_aiter_mla_sparse.py | 1 + 2 files changed, 65 insertions(+) diff --git a/tests/kernels/attention/test_rocm_triton_attn_dsv4.py b/tests/kernels/attention/test_rocm_triton_attn_dsv4.py index 77e068ab17..6fe2a3e775 100644 --- a/tests/kernels/attention/test_rocm_triton_attn_dsv4.py +++ b/tests/kernels/attention/test_rocm_triton_attn_dsv4.py @@ -1,6 +1,8 @@ # SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project +from types import SimpleNamespace + import pytest import torch @@ -196,6 +198,68 @@ def _ragged_from_rows( ) +@torch.inference_mode() +def test_paged_mqa_logits_do_not_contain_nan(monkeypatch) -> None: + from vllm._aiter_ops import rocm_aiter_ops + from vllm.v1.attention.ops import rocm_aiter_mla_sparse as mod + + device = torch.device("cuda") + + class FakeWorkspaceManager: + def get_simultaneous(self, *shapes_and_dtypes): + return [ + torch.empty(shape, dtype=dtype, device=device) + for shape, dtype in shapes_and_dtypes + ] + + def fake_paged_mqa_logits( + q_fp8, + kv_cache_fp8, + weights, + out_logits, + context_lens, + block_tables, + max_seq_len, + **kwargs, + ): + del ( + q_fp8, + kv_cache_fp8, + weights, + context_lens, + block_tables, + max_seq_len, + kwargs, + ) + out_logits.fill_(float("nan")) + + monkeypatch.setattr(mod, "_ON_GFX942", False) + monkeypatch.setattr(mod, "_ON_GFX950", True) + monkeypatch.setattr(rocm_aiter_ops, "is_enabled", lambda: True) + monkeypatch.setattr( + mod, + "paged_mqa_logits_module", + lambda: SimpleNamespace(deepgemm_fp8_paged_mqa_logits=fake_paged_mqa_logits), + ) + monkeypatch.setattr( + mod, "current_workspace_manager", lambda: FakeWorkspaceManager() + ) + + q_fp8 = torch.empty((1, 1, 1, 1), dtype=torch.uint8, device=device) + kv_cache_fp8 = torch.empty((1, 1, 1, 5), dtype=torch.uint8, device=device) + logits = mod.rocm_fp8_paged_mqa_logits( + q_fp8, + kv_cache_fp8, + torch.empty((1, 1), dtype=torch.float32, device=device), + torch.ones(1, dtype=torch.int32, device=device), + torch.zeros((1, 1), dtype=torch.int32, device=device), + torch.empty(0, dtype=torch.int32, device=device), + 1, + ) + + assert not torch.isnan(logits).any() + + @torch.inference_mode() def test_compute_global_topk_ragged_indices_and_indptr() -> None: from vllm.models.deepseek_v4.amd.rocm import ( diff --git a/vllm/v1/attention/ops/rocm_aiter_mla_sparse.py b/vllm/v1/attention/ops/rocm_aiter_mla_sparse.py index 6a28324208..63bad3cfad 100644 --- a/vllm/v1/attention/ops/rocm_aiter_mla_sparse.py +++ b/vllm/v1/attention/ops/rocm_aiter_mla_sparse.py @@ -442,6 +442,7 @@ def rocm_fp8_paged_mqa_logits( KVBlockSize=block_size, WavePerEU=2, ) + out_logits.nan_to_num_(float("-inf")) return out_logits deepgemm_fp8_paged_mqa_logits_stage1 = ( aiter_paged_mqa_logits_module.deepgemm_fp8_paged_mqa_logits_stage1