Log data-loader wait time as an input-starvation metric#564
Merged
Conversation
The schedule runner already timed how long it blocked on the data loader but only warned past a threshold and discarded the value. Accumulate it into the step metrics as `data_wait_time_ms` (0 when never starved), so input starvation is visible in the training logs and W&B. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…g metrics The metric is pre-seeded to 0.0 on logging steps, so the accumulation no longer needs a `.get` fallback. Format it as `.2f ms` in the training log line to match step_time_ms / step_time_average_ms. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
jlamypoirier
added a commit
that referenced
this pull request
Jul 16, 2026
Brings main (#562/#564/#567) into the branch. The one semantic conflict was policy-gradient metrics: main #562 added reward/staleness data metrics (`_register_data_metrics`) to the standalone GRPO/GSPO `_forward_backward`, while this branch reworked registration to compute policy metrics in the fused kernel and register them through `register_combinable_extras`. Kept the fused-metrics design and grafted the data metrics onto every registration path: - `register_combinable_extras` gains `split_index` (base + GRPO/GSPO overrides + 3 call sites) so the compiled and monolithic paths can register data metrics. - GRPO triton-standalone and both `register_combinable_extras` overrides now call `_register_data_metrics`; GSPO standalone already did (auto-merged). - `_policy_metric_definitions` folds in `_data_metric_definitions()`. `test_lm_head` now supplies reward/model_version/documents_seen and asserts the data-metric family via a `reference_data_metrics` helper. CPU validation: test_lm_head + test_lm_losses = 722 passed, 75 skipped (triton/GPU-only skipped locally). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Claude Opus 4.8 note: extracted from PR #553 (RL diagnostics umbrella).
What
The schedule runner already times how long it blocks on the data loader, but only warned past a threshold and discarded the value. This accumulates it into the step metrics as
data_wait_time_ms(0 when never starved), so input starvation is visible in the training logs and W&B.Details
0.0on logging steps so "no wait" reads as0rather than a gap in the series.run_step's returnedtrain_metrics, already spread into the trainer's per-step metrics dict.Single-file, 6-line change to
runner.py. This is the last piece of #553; the other parts landed as #555, #559, #560, #561, #562.🤖 Generated with Claude Code