Repository navigation
Conversation
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
This was referenced Sep 30, 2026
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Phase 2B remains in progress. This PR validates the synchronous prototype. Shared Phase 1 contract integration, the general VIME parameter-conversion/export entry point, and early-exit/PD/speculative compatibility remain incomplete. See the acceptance status.
Synchronous RL rollouts can now opt into the existing recurrent CUDA graphs. Full policy publication copies weights in place, so captured parameter addresses remain valid; decode work counts include executed replay rows and exclude capture warmup.
Stacked on #1 (
feat/ouro-synchronous-rl) for the VIME work in ThinkFlowLab#70. The diff contains this graph integration and its validation.Validation on one Thor with full Ouro-1.4B, BF16, Triton; five alternating-order measured pairs per K after warmup:
All 15 pairs match token IDs, logprobs and block-token counts. Generation after full weight publication matches too. The CUDA unittest passes K changes, policy versions, stable parameter addresses, graph reuse and cache release; Ruff passes.
Companion VIME #5 full-model GRPO completes five updates with fresh-process resume and exact parity across all 272 checkpoint tensors and Adam step metadata. Timing protocol and raw results are committed. Fixed-depth synchronous RL is the validated scope; early-exit, PD and speculative RL are separate work.