fix chunked local attn

Signed-off-by: Benjamin Chislett <[email protected]>

Author: benchislett

vllm/attention/layers/chunked_local_attention.py

@@ -1,7 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import functools
-from typing import ClassVar
 
 import torch
 
@@ -12,11 +11,16 @@
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.v1.attention.backends.utils import (
     AttentionCGSupport,
+    AttentionMetadataBuilder,
     CommonAttentionMetadata,
     make_local_attention_virtual_batches,
     subclass_attention_backend,
 )
-from vllm.v1.kv_cache_interface import ChunkedLocalAttentionSpec, KVCacheSpec
+from vllm.v1.kv_cache_interface import (
+    AttentionSpec,
+    ChunkedLocalAttentionSpec,
+    KVCacheSpec,
+)
 
 from ..layer import Attention
 
@@ -30,9 +34,18 @@ def create_chunked_local_attention_backend(
     prefix = f"ChunkedLocalAttention_{attention_chunk_size}_{block_size}_"
 
     underlying_builder = underlying_attn_backend.get_builder_cls()
+    assert issubclass(underlying_builder, AttentionMetadataBuilder)
 
     class ChunkedLocalAttentionBuilder(underlying_builder):  # type: ignore
-        _cudagraph_support: ClassVar[AttentionCGSupport] = AttentionCGSupport.NEVER
+        @classmethod
+        def get_cudagraph_support(
+            cls: type["AttentionMetadataBuilder"],
+            vllm_config: VllmConfig,
+            kv_cache_spec: AttentionSpec,
+        ) -> AttentionCGSupport:
+            # Explicit override in case the underlying builder specialized this getter.
+            # @override omitted only because of mypy limitation due to type variable.
+            return AttentionCGSupport.NEVER
 
         def build(
             self,