8000 2.2.0+ regresses SDPA performance on Windows · Issue #125070 · pytorch/pytorch · GitHub
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content
2.2.0+ regresses SDPA performance on Windows  #125070
New issue
New issue
Open
Open
2.2.0+ regresses SDPA performance on Windows #125070
Assignees
taras-janea
Labels
module: sdpaAll things related to torch.nn.functional.scaled_dot_product_attentiionmodule: windowsWindows support for PyTorchtriagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate module
@Xemorr

Description

@Xemorr
Xemorr
opened on Apr 26, 2024

🐛 Describe the bug

In PyTorch <2.2.0, SDPA (scaled_dot_product_attention) supports Flash Attention v1 on Windows. In PyTorch 2.2.0>=, it does not support any Flash Attention on Windows.

Versions

This is a report of a regression between 2.1.2 and 2.2.0+

cc @peterjc123 @mszhanyi @skyline75489 @nbcsm @vladimir-aubrecht @iremyux @Blackhex @cristianPanaite

Metadata

Metadata

Assignees

Labels

module: sdpaAll things related to torch.nn.functional.scaled_dot_product_attentiionmodule: windowsWindows support for PyTorchtriagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

Type

No type

Projects

PyTorch On Windows

Status

Blocked

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions
    0