Move client-supplied per-token RL tensors to GPU in get_rollout_data - #7
Merged
Conversation
…vantages/values/returns) to GPU in get_rollout_data; native path computes these on-actor, seam path ships them CPU and loss math crossed devices
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The native path computes log_probs/ref_log_probs/advantages/values/returns on-actor (already on device); the tinker seam ships them from the client as CPU tensors, and loss math crosses devices at ppo_kl = old_log_probs - log_probs (RuntimeError: cuda:1 vs cpu). Move them alongside tokens/loss_masks in get_rollout_data; no-op when already on device (native path unaffected).
Validated: GSM8K RL recipe end-to-end on the pooled backend (12 rounds, kl_sample_train 0.003-0.02 nats), plus 2 SFT + 2 RL mixed-tenant run; SFT gates unaffected.