$ the-wire · showcase
vLLM rollouts stop hiding truncation, MILES reworks sampling and grading
By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology
verl now propagates vLLM's truncation reason through to clients, while MILES gains priority ordering inside its uniform and exposure pools and a native Nemotron abstention reward.
[vllm, rollout] fix: preserve the length stop reason in token outputs verl-project/verl
vLLM's finish_reason="length" now survives into TokenOutput.stop_reason, where it used to arrive as completed for both a natural stop and an exhausted output budget. If you translate that value to the Anthropic Messages API, a truncated response no longer comes back labeled end_turn, which is what let clients exit with work still in progress.
Compose priority ordering with uniform and exposure pool sampling allenai/open-instruct
You can now set data.sampling.priority.field as a modifier on the uniform and exposure providers, so allocation still picks a pool but that pool consumes priority groups 1 through 4 with seeded shuffling. Previously exposure balancing and priority-first ordering were separate provider choices, so you could not prefer screened prompts while balancing trained exposure across sources.
[rollout, vllm, sglang] fix: launch rollout servers when workers hold no accelerator; per-replica env-var hook verl-project/verl
vLLMReplica.launch_servers demanded an accelerator id from every rollout worker via get_accelerator_ids() and SGLangReplica read visible_devices_keyword from the environment, so both failed on any worker holding no accelerator, TPU included. Out-of-tree platforms can now reuse launch_servers for vLLM and SGLang instead of copying it, and per-replica env-var overrides are settable.
Support native Nemotron abstention QA rewards in MILES allenai/open-instruct
The nemo_abstention_v1 adapter preserves system and user inputs and emits structured question/reference fields for Nemotron-RL-QA-Abstention-v1, paired with NeMo Gym's three-tier judged reward. Ordinary answer matching scored an abstention identically to a wrong answer, so a policy trained on it had no signal to abstain when uncertain.
Unify MILES grading inputs with typed reward components allenai/open-instruct
Answer components receive the presented response when final-answer grading is enabled, while process and format components get their own declared context, replacing verifiers that each picked different parts of the same response. One shared representation now identifies thinking, the presented answer or action, and completion state.