fix(attention): restore MLA DCP query state

aphrodite/model_executor/layers/attention/mla_attention.py

@@ -767,11 +767,14 @@ def forward_impl(
         num_mha_tokens = q.size(0) - num_mqa_tokens
 
         if self.impl.is_sparse and num_mha_tokens > 0:
+            impl_forward_mha = getattr(type(self.impl), "forward_mha", None)
+            mha_available = impl_forward_mha is not MLAAttentionImpl.forward_mha
             prefill = getattr(attn_metadata, "prefill", None)
-            use_dense_mha = getattr(prefill, "use_dense_mha", False)
+            use_dense_mha = mha_available and getattr(prefill, "use_dense_mha", False)
             prefill_max_seq_len = attn_metadata.prefill_max_seq_len  # type: ignore[attr-defined]
             use_masked_mha = (
-                self.prefill_backend is not None
+                mha_available
+                and self.prefill_backend is not None
                 and self.impl.masked_mha_available  # type: ignore[attr-defined]
                 and self.impl.dcp_world_size <= 1
                 and prefill is not None

aphrodite/model_executor/layers/mla.py

@@ -98,6 +98,10 @@ def __init__(
         # the topk_tokens buffer written by a previous layer in the same pass.
         # Refer: https://arxiv.org/abs/2603.12201 for more details.
         self.skip_topk = skip_topk
+        # qrep is active when the query projection is a DCP-group-sharded layer
+        # that materializes the full group head set locally.
+        q_proj_layer = self.q_b_proj if self.q_lora_rank is not None else self.q_proj
+        self.dcp_q_replicate = getattr(q_proj_layer, "qrep_active", False)
         if self.indexer is not None:
             assert hasattr(self.indexer, "topk_tokens")
             self.topk_tokens = self.indexer.topk_tokens
@@ -115,6 +119,7 @@ def __init__(
             quant_config=quant_config,
             prefix=f"{prefix}.attn",
             kv_b_proj=self.kv_b_proj,
+            dcp_q_replicate=self.dcp_q_replicate,
             use_sparse=self.is_sparse,
             indexer=self.indexer,
             topk_indices_buffer=mla_modules.topk_indices_buffer,

aphrodite/platforms/cuda.py

@@ -694,10 +694,12 @@ def support_static_graph_mode(cls) -> bool:
 
     @classmethod
     def support_deep_gemm(cls) -> bool:
-        """Currently, only Hopper and Blackwell GPUs are supported."""
-        return (
-            cls.is_device_capability(90) or cls.is_device_capability_family(100) or cls.is_device_capability_family(120)
-        )
+        """Return whether DeepGEMM is validated on this CUDA architecture.
+
+        Consumer Blackwell (SM12x) is excluded. The available DeepGEMM
+        revisions either lack its kernels or produce incorrect results there.
+        """
+        return cls.is_device_capability(90) or cls.is_device_capability_family(100)
 
     @classmethod
     def is_integrated_gpu(cls, device_id: int = 0) -> bool:

aphrodite/v1/attention/backend.py

@@ -972,6 +972,10 @@ class MLAAttentionImpl(AttentionImplBase[T], Generic[T]):
     """MLA attention implementation with forward_mqa and forward_mha methods."""
 
     supports_pcp: bool = True
+    # Masked sparse-MHA prefill is an optional capability implemented by
+    # SparseMLACommonImpl. Platform-specific sparse MLA backends inherit the
+    # safe default and continue to route prefills through MQA.
+    masked_mha_available: bool = False
 
     @abstractmethod
     def __init__(