Add Dr. GRPO loss normalization - #853
Open
zjn20030811 wants to merge 1 commit into
Open
Conversation
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Addresses #364.
Threads TRL-compatible loss_type values (grpo, bnpo, dr_grpo) from public trainer configuration through local, serverless, Unsloth, and Megatron paths. Dr. GRPO uses a fixed B * max_completion_length denominator, including packed completion tails and context-parallel ownership. Incomplete accumulation padding is marked explicitly so dummy microbatches contribute zero while real empty completions retain their batch semantics.
The implicit completion length follows TRL 0.20s 256-token default. Regression tests cover unequal lengths, packed tails, masks, zero-token batches, dummy padding, configuration plumbing, and context-parallel denominator ownership.
Checks: py -3.12 -m pytest -q tests/unit/test_loss_reduction.py src/art/test/test_kl_advantage.py (20 passed); ruff check; ruff format --check; py -3.12 -m compileall. Broader backend suites require Linux/GPU-only dependencies unavailable in the local Windows environment.