Repository navigation
fix(ppo): promote reward shaping to at least FP32 - #1334
Merged
Merged
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Grouped advantage estimation currently inherits the reward tensor's dtype. A BF16 reward model can therefore round away the group baseline: rewards
[1, 1, 1, 1.0078125]produce approximately[0, 0, 0, 2]for group-normalized advantages instead of[-0.5, -0.5, -0.5, 1.5], changing the PPO gradient. Integer rewards also fail group statistics or fractional penalties, and constant FP16 groups can produce NaN.Promote rewards to at least FP32 (preserving FP64 inputs) before the existing length-penalty and group-statistics code. This keeps reward processing in a stable compute dtype for all sources and reuses the existing synchronization of logged rewards.
Validation: full pytest suite (92 passed), compileall and pre-commit. Tests cover six estimators, reward dtypes including FP64, constant groups, fractional penalties, and a reference PPO gradient. A CPU probe with the actual OpenRLHF reward head on a randomly initialized tiny BF16 Llama and recombined SQuAD/CMRC/HumanEval texts also matches a float64 arithmetic reference after the fix. This is numerical validation, not a pretrained-model quality benchmark or GPU training run.