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 11, 10:00 UTC
[Bugfix] Fix broken ViT attention selection for Blackwell device (#30731)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Isotr0py committed
Dec 16, 2025 at 13:24 UTC
e94384bbadbaf99dea24c4af4de6a8c897f830e7
1 file changed
+2
-8
vllm/model_executor/models/vision.py
+2
-8
@@ -11,7 +11,7 @@ import torch
11
from transformers import PretrainedConfig
12
13
from vllm.attention.backends.registry import AttentionBackendEnum
14
-from vllm.config import VllmConfig, get_current_vllm_config
14
+from vllm.config import VllmConfig
15
from vllm.distributed import (
16
get_tensor_model_parallel_rank,
17
get_tensor_model_parallel_world_size,
@@ -88,16 +88,10 @@ def get_vit_attn_backend(
88
"""
89
Get the available attention backend for Vision Transformer.
90
"""
91
- attn_backend = attn_backend_override
92
-
93
- selected_backend = get_current_vllm_config().attention_config.backend
94
- if attn_backend is None:
95
- attn_backend = selected_backend
96
-
91
return current_platform.get_vit_attn_backend(
92
head_size,
93
dtype,
100
- backend=attn_backend,
94
+ backend=attn_backend_override,
95
)
96
97