biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix][ROCm] Fix Unsupported attention metadata type for speculative decoding in eagle.py (#31714)

Browse Source 
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
This commit is contained in:
vllmellm
2026-01-06 08:53:22 +01:00
committed by GitHub
parent da71d44410
commit e9717801bd
1 changed files with 6 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/v1/spec_decode/eagle.py
View File

@@ -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)