Description

This PR adds caching for multiple attention configurations in the same run.

Currently we only cache one attention_params and its corresponding available_backends and selected_backend. If there are multiple configurations in the same training/inference job, for example, it runs like this: config 1, config 2, ..., config 1, config 2, ..., then the second config 1 or config 2 still needs to go through the get_attention_backend() call, which is CPU intensive. With this PR, we cache the backend analysis results for multiple configs, so the second time a config is run, no repeated analysis is needed. The max number of configs cached is currently set to 10.

Fixes #1349

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refractor

Changes

Please list the changes introduced in this PR:

• Add caching for attention backend selection results

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes