[BugFix] ChunkedLocalAttention is currently not CG compatible (vllm-project#26034)

Signed-off-by: Lucas Wilkinson <[email protected]>

Author: LucasWilkinson

vllm/attention/layers/chunked_local_attention.py

@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import functools
-from typing import List, Optional
+from typing import ClassVar, List, Optional
 
 import torch
 
@@ -12,8 +12,8 @@
 from vllm.config import CacheConfig
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.v1.attention.backends.utils import (
-    CommonAttentionMetadata, make_local_attention_virtual_batches,
-    subclass_attention_backend)
+    AttentionCGSupport, CommonAttentionMetadata,
+    make_local_attention_virtual_batches, subclass_attention_backend)
 
 from ..layer import Attention
 
@@ -29,6 +29,8 @@ def create_chunked_local_attention_backend(
     underlying_builder = underlying_attn_backend.get_builder_cls()
 
     class ChunkedLocalAttentionBuilder(underlying_builder):  # type: ignore
+        cudagraph_support: ClassVar[AttentionCGSupport] = \
+            AttentionCGSupport.NEVER
 
         def build(self,
                   common_prefix_len: int,