Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Misc] Set default backend to SDPA for get_vit_attn_backend #12235

Merged
merged 1 commit into from
Jan 21, 2025
Merged

[Misc] Set default backend to SDPA for get_vit_attn_backend #12235

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
30 changes: 16 additions & 14 deletions vllm/model_executor/models/vision.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -82,23 +82,25 @@ def get_vit_attn_backend(support_fa: bool = False) -> _Backend:
if backend_by_env_var is not None:
selected_backend = backend_name_to_enum(backend_by_env_var)
if selected_backend is None:
# For Volta and Turing GPUs, use xformers instead.
device_available = current_platform.has_device_capability(80)
if device_available and support_fa:
from transformers.utils import is_flash_attn_2_available
if is_flash_attn_2_available():
selected_backend = _Backend.FLASH_ATTN
if current_platform.is_cuda():
device_available = current_platform.has_device_capability(80)
if device_available and support_fa:
from transformers.utils import is_flash_attn_2_available
if is_flash_attn_2_available():
selected_backend = _Backend.FLASH_ATTN
else:
logger.warning_once(
"Current `vllm-flash-attn` has a bug inside vision "
"module, so we use xformers backend instead. You can "
"run `pip install flash-attn` to use flash-attention "
"backend.")
selected_backend = _Backend.XFORMERS
else:
logger.warning_once(
"Current `vllm-flash-attn` has a bug inside vision module, "
"so we use xformers backend instead. You can run "
"`pip install flash-attn` to use flash-attention backend.")
# For Volta and Turing GPUs, use xformers instead.
selected_backend = _Backend.XFORMERS
elif current_platform.is_cpu() or current_platform.is_rocm():
# ROCM doesn't support xformers
selected_backend = _Backend.TORCH_SDPA
else:
selected_backend = _Backend.XFORMERS
# Default to torch SDPA for other non-GPU platforms.
selected_backend = _Backend.TORCH_SDPA
return selected_backend


Expand Down
Loading