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

Remove upstream fa checks (#29471)

Browse Source 
Signed-off-by: mingyuanm <mingyuanm@nvidia.com>
Signed-off-by: Roger Wang <hey@rogerw.io>
Co-authored-by: Roger Wang <hey@rogerw.io>
This commit is contained in:
Mingyuan Ma
2025-11-28 05:52:42 -08:00
committed by GitHub
parent e2f56c309d
commit 460d8bbf2d
13 changed files with 18 additions and 148 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/model_executor/models/glm4_1v.py
View File

@@ -47,10 +47,7 @@ from transformers.models.glm4v.video_processing_glm4v import Glm4vVideoProcessor
from transformers.video_utils import VideoMetadata
from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.attention.layer import (
check_upstream_fa_availability,
maybe_get_vit_flash_attn_backend,
)
from vllm.attention.layer import maybe_get_vit_flash_attn_backend
from vllm.config import VllmConfig
from vllm.config.multimodal import BaseDummyOptions, VideoDummyOptions
from vllm.distributed import get_tensor_model_parallel_world_size, parallel_state
@@ -296,12 +293,10 @@ class Glm4vVisionAttention(nn.Module):
dtype=torch.get_default_dtype(),
attn_backend_override=attn_backend_override,
)
self.use_upstream_fa = False
self.attn_backend, self.flash_attn_varlen_func = (
maybe_get_vit_flash_attn_backend(
self.attn_backend,
self.use_upstream_fa,
attn_backend_override=attn_backend_override,
)
)
@@ -730,11 +725,6 @@ class Glm4vVisionTransformer(nn.Module):
dtype=torch.get_default_dtype(),
attn_backend_override=attn_backend_override,
)
if (
self.attn_backend != AttentionBackendEnum.FLASH_ATTN
and check_upstream_fa_availability(torch.get_default_dtype())
):
self.attn_backend = AttentionBackendEnum.FLASH_ATTN
@property
def dtype(self) -> torch.dtype: