Skip to content

Conversation

@yzh119
Copy link
Member

@yzh119 yzh119 commented Mar 12, 2023

Motivation

Currently, some of the schedule checks are too strict, which makes it hard to schedule some workloads such as FlashAttention whose reduction is two-stage and does not strictly follows our standard.

This PR adds an optional argument disable_checks which mutes some checks. The argument defaults to False and we can enable it whenever we want to disable some soft checks (by soft we mean if we violate such checks, the schedule is not necessarily invalid, and if we violate hard checks the schedule step is invalid).

In the future, we should collect the soft and hard checks for all schedule primitives. This PR serves for FlashAttention and only cares bind and some reduction primitives for now.

@spectrometerHBH @junrushao @Hzfengsy

@tvm-bot
Copy link
Collaborator

tvm-bot commented Mar 12, 2023

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

  • No users to tag found in teams: schedule See #10317 for details

Generated by tvm-bot

@tqchen
Copy link
Member

tqchen commented Mar 19, 2023

@tvm-bot rerun

Copy link
Member

@junrushao junrushao left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@junrushao junrushao merged commit 542274d into apache:main Mar 19, 2023
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

5 participants