diff --git a/tests/v1/attention/test_mla_prefill_registry.py b/tests/v1/attention/test_mla_prefill_registry.py index dfa3a029ce..52c8d18554 100644 --- a/tests/v1/attention/test_mla_prefill_registry.py +++ b/tests/v1/attention/test_mla_prefill_registry.py @@ -28,6 +28,28 @@ class CustomMLAPrefillBackend(MLAPrefillBackend): raise NotImplementedError +def test_prefill_backend_clone_has_isolated_metadata(): + backend = CustomMLAPrefillBackend( + num_heads=4, + scale=0.5, + kv_lora_rank=8, + qk_nope_head_dim=16, + qk_rope_head_dim=8, + v_head_dim=32, + vllm_config=object(), + ) + + clone = backend.clone() + + assert isinstance(clone, CustomMLAPrefillBackend) + assert clone is not backend + assert clone.num_heads == backend.num_heads + assert clone.scale == backend.scale + backend._prefill_metadata = object() + clone._prefill_metadata = object() + assert clone._prefill_metadata is not backend._prefill_metadata + + @pytest.fixture(autouse=True) def cleanup_overrides(): """Clear any overrides after each test.""" diff --git a/vllm/model_executor/layers/attention/mla_attention.py b/vllm/model_executor/layers/attention/mla_attention.py index e1cb20d2e7..027f19a90b 100644 --- a/vllm/model_executor/layers/attention/mla_attention.py +++ b/vllm/model_executor/layers/attention/mla_attention.py @@ -1563,9 +1563,12 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]): device=device, ) + # Metadata builders are created per ubatch when DBO is enabled. MLA + # prefill backends keep the prepared metadata on the backend object, so + # each builder needs its own backend instance to avoid cross-ubatch races. self._prefill_backend = self.compilation_config.static_forward_context[ layer_names[0] - ].prefill_backend + ].prefill_backend.clone() supports_spec_decode = self.query_len_support != QueryLenSupport.SINGLE_ONLY self._init_reorder_batch_threshold( diff --git a/vllm/v1/attention/backends/mla/prefill/base.py b/vllm/v1/attention/backends/mla/prefill/base.py index ff478aec4a..c56f3d46c7 100644 --- a/vllm/v1/attention/backends/mla/prefill/base.py +++ b/vllm/v1/attention/backends/mla/prefill/base.py @@ -116,6 +116,17 @@ class MLAPrefillBackend(ABC): self.v_head_dim = v_head_dim self.vllm_config = vllm_config + def clone(self) -> "MLAPrefillBackend": + return self.__class__( + num_heads=self.num_heads, + scale=self.scale, + kv_lora_rank=self.kv_lora_rank, + qk_nope_head_dim=self.qk_nope_head_dim, + qk_rope_head_dim=self.qk_rope_head_dim, + v_head_dim=self.v_head_dim, + vllm_config=self.vllm_config, + ) + def prepare_metadata( # noqa: B027 self, prefill_metadata: "MLACommonPrefillMetadata",