Skip to content

Conversation

@dhernandez0
Copy link
Contributor

@dhernandez0 dhernandez0 commented Dec 3, 2025

Motivation

To avoid really long tuning time, we can set the default value of these params to false. I think if migraphx didn't provide those, it's a safe assumption that they are false.

Please, review MITuna PR as well: ROCm/MITuna#1015

Technical Details

Just set the value to false if they are not in the problem config.

Test Plan

All tests pass.

Test Result

All tests pass.

Submission Checklist

@dhernandez0 dhernandez0 requested a review from causten as a code owner December 3, 2025 13:31
@dhernandez0 dhernandez0 self-assigned this Dec 3, 2025
Copilot finished reviewing on behalf of dhernandez0 December 3, 2025 13:35
Copy link
Contributor

Copilot AI left a comment

Choose a reason for hiding this comment

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

Pull request overview

This PR modifies the attention configuration generation to reduce tuning time by setting default values for causal, return_lse, with-attn-scale, and with-attn-bias parameters to false when they are not explicitly specified in the problem configuration. This is a performance optimization that reduces the configuration search space by avoiding unnecessary testing of these feature combinations.

Key Changes

  • Changed four attention parameters in the default test space from bool_space (['false', 'true']) to default_to_false (['false']), effectively reducing the number of generated test configurations
  • Added a comment explaining that undefined parameters should default to false
  • This aligns with existing defaults in AttentionConfiguration.from_command_line() and matches common usage patterns seen in tier1-attention-configs

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

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.

4 participants