A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 10, 16:30 UTC
[Bugfix][ROCm] Fix Unsupported attention metadata type for speculative decoding in `eagle.py` (#31714)
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
vllmellm committed
Jan 6, 2026 at 08:53 UTC
e9717801bdc5f8dec45ee1d1a8cebed9f4a13544
1 file changed
+6
-2
vllm/v1/spec_decode/eagle.py
+6
-2
@@ -27,7 +27,6 @@ from vllm.multimodal import MULTIMODAL_REGISTRY
27
from vllm.platforms import current_platform
28
from vllm.triton_utils import triton
29
from vllm.utils.platform_utils import is_pin_memory_available
30
-from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
30
from vllm.v1.attention.backends.tree_attn import (
31
TreeAttentionMetadata,
32
TreeAttentionMetadataBuilder,
@@ -167,7 +166,12 @@ class EagleProposer:
166
# Determine allowed attention backends once during initialization.
167
self.allowed_attn_types: tuple | None = None
168
if current_platform.is_rocm():
170
- rocm_types = [TritonAttentionMetadata, FlashAttentionMetadata]
169
+ from vllm.v1.attention.backends.rocm_attn import RocmAttentionMetadata
170
+
171
+ rocm_types = [
172
+ TritonAttentionMetadata,
173
+ RocmAttentionMetadata,
174
+ ]
175
# ROCM_AITER_FA is an optional backend
176
if find_spec(
177
AttentionBackendEnum.ROCM_AITER_FA.get_path(include_classname=False)