Skip to content

[Feature] Log efficient_attn_ratio and img_efficient_attn_ratio to exp_tracker#1666

Open
claude[bot] wants to merge 1 commit intomainfrom
feat/log-efficient-attn-ratio
Open

[Feature] Log efficient_attn_ratio and img_efficient_attn_ratio to exp_tracker#1666
claude[bot] wants to merge 1 commit intomainfrom
feat/log-efficient-attn-ratio

Conversation

@claude
Copy link
Copy Markdown
Contributor

@claude claude bot commented Apr 10, 2026

Summary

  • Add efficient_attn_ratio and img_efficient_attn_ratio to the experiment tracker scalars under the runtime_info/ prefix in Trainer._log_step()
  • Update FakeEngine mock in tests to include the missing step_consumed_img_tokens and img_efficient_attn_ratio fields

Changes

  • xtuner/v1/train/trainer.py: Added two new entries to log_scalars dict
  • tests/train/test_trainer.py: Updated FakeEngine.train_step() return value to match full TrainStepInfo schema

Test plan

  • Existing trainer tests should continue to pass with the updated mock
  • The new scalar entries will appear in experiment tracking dashboards (TensorBoard, WandB, etc.)

Closes #1665

…p_tracker

Add efficient_attn_ratio and img_efficient_attn_ratio to the log_scalars
dict in _log_step so they are recorded by the experiment tracker under the
runtime_info/ prefix. Also update FakeEngine mock in tests to include the
missing step_consumed_img_tokens and img_efficient_attn_ratio fields.

Closes #1665
@HAOCHENYE
Copy link
Copy Markdown
Collaborator

@claude directly merge it after unit test passed

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.

[Feature Requst] Update more training info into exp_tracking

1 participant