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

[Bugfix] Fallback ViT attn backend to SDPA for blackwell (#25851)

Browse Source 
Signed-off-by: Roger Wang <hey@rogerw.io>
This commit is contained in:
Roger Wang
2025-09-28 23:03:51 -07:00
committed by GitHub
parent 143844fa43
commit 65ecb4f134
2 changed files with 7 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
vllm/model_executor/models/qwen3_vl.py
View File

@@ -66,7 +66,7 @@ from vllm.multimodal.processing import (BaseMultiModalProcessor,
PromptReplacement, PromptUpdate,
PromptUpdateDetails)
from vllm.multimodal.profiling import BaseDummyInputsBuilder
from vllm.platforms import _Backend, current_platform
from vllm.platforms import _Backend
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.config import uses_mrope
from vllm.utils import is_list_of
@@ -336,14 +336,6 @@ class Qwen3_VisionTransformer(nn.Module):
}:
raise RuntimeError(
f"Qwen3-VL does not support {self.attn_backend} backend now.")
if current_platform.is_device_capability(
100) and self.attn_backend != _Backend.TORCH_SDPA:
# TODO(Roger/Wentao): remove this after FA
# or XFORMERS's issue fixed on Blackwell
logger.info_once("Qwen3-VL vision attention does not support "
f"{self.attn_backend} backend on Blackwell now. "
"Vision attention backend is set to TORCH_SDPA.")
self.attn_backend = _Backend.TORCH_SDPA
self.blocks = nn.ModuleList([
Qwen3_VisionBlock(