Repository navigation
agent: leave out of each infer the observations the server holds - reuse-feedback-obs, SPEC 10.1 - #16
Merged
Merged
Conversation
…nReplaced); probe-v1 env for plugrl-conformance --probe The probe checker's resync scenario failed this client: after the agent reconnected inside its next call, envs mid-chunk kept executing chunks from the old connection and sent their feedback on the new one, and a failed feedback was followed by a reconnect inside the next feedback call. SPEC 7.6 forbids both. The agent now raises ConnectionReplaced instead, feedback never reconnects, and the rollout drops every chunk in flight and re-plans every env; ended episodes are still reset. The reconnect-on-server-stop resend (the SPEC 7.6 Gap) is gone with it.
…C 10.1, reuse-feedback-obs) A feedback carries the observation the next chunk starts from, and the next infer sent it again. When the server lists reuse-feedback-obs in its metadata, the agent now marks in `reuse` each row the server already has and sends only the others. The decision is the agent's, not the rollout's, so no caller has to know the feature exists. A row is left out only when: - the server lists the feature and --reuse-feedback-obs is on (the default); - the env's last feedback on this connection did not end its episode, since that feedback carried the terminal observation and not the reset one; - the row is identical to the one that feedback carried. The agent keeps its own copy, so a caller that reuses its buffers cannot make it skip a row that changed. A new connection holds nothing, so the record is cleared on every close and reconnect. plugrl-protocol is pinned to the commit of plugrl-protocol#11 that adds the constant; it moves to main once that merges.
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 client side of plugrl-protocol#11. A feedback carries the observation the next chunk starts from, and the next infer used to send it again. When the server lists
reuse-feedback-obs, the agent now leaves those rows out of the infer and marks them inreuse. Stacked on #15: merge that first. Only the last commit is new.The decision is the agent's, so neither the rollout nor any other caller changes. A row is left out only when all of these hold:
--reuse-feedback-obsis on (the default;--no-reuse-feedback-obsturns it off);A new connection holds nothing, so the record is cleared on every close and reconnect.
Checks:
tests/test_reuse_feedback_obs.py, 7 cases:probe-v1:--features reuse-feedback-obs: no violations. 9 infers reused rows, all of them rows the server held, including after the resync.reusesent.plugrl-protocolis pinned to the commit of plugrl-protocol#11 that addsREUSE_FEEDBACK_OBS. It moves to main once that merges.