Do not abort in-flight requests when bumping the engine weight version

[yueming-yuan]

Problem

Every weight update ends by aborting all in-flight rollout requests. The chain:

1. _finalize_and_resume_engines (mixin.py) calls engine.update_weight_version(new_version) — added so full-param and LoRA updates share one version bump. The July-era code passed weight_version inside the update_weights_from_distributed payload instead and never hit this endpoint.
2. The wrapper posts only {"new_version": ...} (sglang_engine.py).
3. sglang's UpdateWeightVersionReqInput defaults abort_all_requests: bool = True, and the /update_weight_version handler then runs abort_request(abort_all=True) — one step before continue_generation.

Under pause_generation_mode=in_place this defeats the pause itself: the scheduler's in-place pause deliberately leaves all running requests intact so they resume decoding under the new weights, and the version bump then destroys them.

Downstream, each aborted turn returns a partial message with finish_reason: "abort". The agent extracts no command from the partial and ends the episode, the sample finishes ABORTED, and check_no_aborted drops the entire group.

Measured impact (16-node GLM-5.2 agentic run, job 2403; July build as control)

• ~65 aborts per update ≈ the number of requests decoding when the update lands; abort timestamps sit 30–60s after each pause (after the ~44s transfer, at finalize), with none during the transfer window.
• 38% of all samples end ABORTED (July control: 0%). 11–18 groups (88–144 samples) discarded per step — about two thirds of rollout production.
• Consumable sample supply drops to ~64 per 450s while raw production is ~178 per step; train_wait_time 137–340s per step.
• The abort_all sweep racing _wait_one_response wake-ups is also the source of the rid_to_state KeyErrors logged during updates (343 in this run).
• Mixed-version turns disappear (0.11–0.56 mixed_version in the July control vs 0–9%): any turn spanning an update is killed instead of finishing under the new weights.

Fix

Pass abort_all_requests: False from the wrapper. The pause mode has already decided the in-flight requests' fate by the time the version is bumped; under in_place they must stay alive. This restores the July behavior, under which a 99-step run trained with turns routinely finishing under mixed weights (the per-turn rollout log-probs record the actual sampling distribution either way, so importance correction is unaffected).

A follow-up candidate (not in this PR): the agent loop in examples/experimental/openenv does not inspect finish_reason, so any residual aborted turn still ends its episode; the session server's retry rollback (prepare_pretokenized) already supports re-issuing such a turn.

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}