[Bugfix][ROCm] Fix Unsupported attention metadata type for speculative...

[Bugfix][ROCm] Fix Unsupported attention metadata type for speculative decoding in `eagle.py` (#31714)
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>

vllm/v1/spec_decode/eagle.py

@@ -27,7 +27,6 @@ from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.platforms import current_platform
 from vllm.triton_utils import triton
 from vllm.utils.platform_utils import is_pin_memory_available
-from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
 from vllm.v1.attention.backends.tree_attn import (
     TreeAttentionMetadata,
     TreeAttentionMetadataBuilder,
@@ -167,7 +166,12 @@ class EagleProposer:
         # Determine allowed attention backends once during initialization.
         self.allowed_attn_types: tuple | None = None
         if current_platform.is_rocm():
-            rocm_types = [TritonAttentionMetadata, FlashAttentionMetadata]
+            from vllm.v1.attention.backends.rocm_attn import RocmAttentionMetadata
+
+            rocm_types = [
+                TritonAttentionMetadata,
+                RocmAttentionMetadata,
+            ]
             # ROCM_AITER_FA is an optional backend
             if find_spec(
                 AttentionBackendEnum.ROCM_AITER_FA.get_path(include_classname=False)