[ROCm][perf] fix Aiter sparse MLA with MTP>1 (#37887)

Signed-off-by: Stig-Arne Grönroos <stig-arne.gronroos@amd.com>
Signed-off-by: Stig-Arne Grönroos <sgronroo@amd.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>

vllm/v1/spec_decode/dflash.py

@@ -257,19 +257,19 @@ class DFlashProposer(SpecDecodeBaseProposer):
         )
 
     @override
-    def build_per_layer_attn_metadata(
+    def build_per_group_and_layer_attn_metadata(
         self, cad: CommonAttentionMetadata, draft_index: int = 0
-    ) -> dict[str, object]:
+    ) -> tuple[list[object], dict[str, object]]:
-        per_layer_attention_metadata = super().build_per_layer_attn_metadata(
+        per_group, per_layer = super().build_per_group_and_layer_attn_metadata(
             cad, draft_index
         )
-        for layer_name, attn_metadata in per_layer_attention_metadata.items():
+        for layer_name, attn_metadata in per_layer.items():
             assert getattr(attn_metadata, "causal", None) is False, (
                 f"Attention metadata for layer {layer_name} does not have"
                 " non-causal support, which is required for DFlash."
                 " Consider using a different attention backend, such as FlashAttention."
             )
-        return per_layer_attention_metadata
+        return per_group, per_layer
 
     @override
     def _get_eagle3_use_aux_hidden_state_from_config(self):

vllm/v1/spec_decode/eagle.py

@@ -225,6 +225,9 @@ class SpecDecodeBaseProposer:
         # Determine allowed attention backends once during initialization.
         self.allowed_attn_types: tuple | None = None
         if current_platform.is_rocm():
+            from vllm.v1.attention.backends.mla.indexer import (
+                DeepseekV32IndexerMetadata,
+            )
             from vllm.v1.attention.backends.mla.rocm_aiter_mla_sparse import (
                 ROCMAiterMLASparseMetadata,
             )
@@ -234,6 +237,7 @@ class SpecDecodeBaseProposer:
                 TritonAttentionMetadata,
                 RocmAttentionMetadata,
                 ROCMAiterMLASparseMetadata,
+                DeepseekV32IndexerMetadata,
             ]
             # ROCM_AITER_FA is an optional backend
             # We check is_enabled() here to avoid importing the backend module during
@@ -444,8 +448,8 @@ class SpecDecodeBaseProposer:
             )
         )
 
-        per_layer_attn_metadata = self.build_per_layer_attn_metadata(
+        per_group_attn_metadata, per_layer_attn_metadata = (
-            common_attn_metadata
+            self.build_per_group_and_layer_attn_metadata(common_attn_metadata)
         )
 
         cudagraph_runtime_mode, num_input_tokens, num_tokens_across_dp = (
@@ -486,10 +490,7 @@ class SpecDecodeBaseProposer:
             positions = self.positions[token_indices_to_sample]
         hidden_states = hidden_states[token_indices_to_sample]
 
-        if any(
+        if any(isinstance(md, TreeAttentionMetadata) for md in per_group_attn_metadata):
-            isinstance(attn_metadata, TreeAttentionMetadata)
-            for attn_metadata in per_layer_attn_metadata.values()
-        ):
             # Draft using tree attention - requires full logits for top-k
             logits = self.model.compute_logits(sample_hidden_states)
             draft_token_ids_list = self.propose_tree(
@@ -505,16 +506,15 @@ class SpecDecodeBaseProposer:
 
         draft_token_ids = self._greedy_sample(sample_hidden_states)
 
-        for attn_metadata in per_layer_attn_metadata.values():
+        if self.allowed_attn_types is not None:
-            if self.allowed_attn_types is not None and not isinstance(
+            for group_md in per_group_attn_metadata:
-                attn_metadata, self.allowed_attn_types
+                if not isinstance(group_md, self.allowed_attn_types):
-            ):
+                    raise ValueError(
-                raise ValueError(
+                        f"Unsupported attention metadata type for speculative "
-                    f"Unsupported attention metadata type for speculative "
+                        "decoding with num_speculative_tokens > 1: "
-                    "decoding with num_speculative_tokens > 1: "
+                        f"{type(group_md)}. Supported types are: "
-                    f"{type(attn_metadata)}. Supported types are: "
+                        f"{self.allowed_attn_types}"
-                    f"{self.allowed_attn_types}"
+                    )
-                )
 
         # Generate the remaining draft tokens.
         draft_token_ids_list = [draft_token_ids]
@@ -595,7 +595,7 @@ class SpecDecodeBaseProposer:
                 common_attn_metadata._num_computed_tokens_cpu += 1
 
             # Rebuild attention metadata
-            per_layer_attn_metadata = self.build_per_layer_attn_metadata(
+            _, per_layer_attn_metadata = self.build_per_group_and_layer_attn_metadata(
                 common_attn_metadata, draft_index=token_index + 1
             )
 
@@ -809,17 +809,19 @@ class SpecDecodeBaseProposer:
 
         return model_kwargs, num_input_tokens
 
-    def build_per_layer_attn_metadata(
+    def build_per_group_and_layer_attn_metadata(
         self, common_attn_metadata: CommonAttentionMetadata, draft_index: int = 0
-    ) -> dict[str, object]:
+    ) -> tuple[list[object], dict[str, object]]:
+        per_group_attn_metadata: list[object] = []
         per_layer_attn_metadata: dict[str, object] = {}
         for attn_group in self.draft_attn_groups:
             attn_metadata = attn_group.get_metadata_builder().build_for_drafting(
                 common_attn_metadata=common_attn_metadata, draft_index=draft_index
             )
+            per_group_attn_metadata.append(attn_metadata)
             for layer_name in attn_group.layer_names:
                 per_layer_attn_metadata[layer_name] = attn_metadata
-        return per_layer_attn_metadata
+        return per_group_attn_metadata, per_layer_attn_metadata
 
     def model_returns_tuple(self) -> bool:
         return self.method not in ("mtp", "draft_model", "dflash")