$ the-wire · showcase
MILES wall-clock stops resume by default, retained generation state in learning examples
By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology
Both MILES pull requests flip defaults that change what a stopped run does and how the learning examples pause generation, while the verl desk's day is almost entirely Ascend CI plumbing.
MILES: resumable wall-clock stop by default (core.max_run_final opts out) allenai/open-instruct
A wall-clock stop on `core.max_run_seconds` used to force a checkpoint, run the end-of-run HF export and final background eval at the stopped update, and mark `workflow.json` complete so a rerun under the same output directory refused to continue. It now resumes by default; set `core.max_run_final = true` to get the old export-and-complete behavior back. If your operators rely on a timed stop m...
MILES: retain generation state by default in learning examples allenai/open-instruct
The one-node and three-node learning examples now keep active-request attention KV and hybrid recurrent state across weight publication instead of retracting and re-prefilling generated prefixes, selected via `miles.pause_generation_mode = "in_place"` and the new `olmo35-b300-12.5@0.8` preset. The optimized B300 pools and CUDA graphs are preserved, but cross-request prefix reuse is disabled, so...
[hardware, vllm, doc] chore: upgrade Ascend vLLM stack to 0.30.0 verl-project/verl
verl's Ascend path moves to the vLLM 0.30.0 stack, so operators running rollout on Ascend NPUs should match that vLLM version alongside this checkout rather than trailing it.
Set 30-minute minimum runtimes for short Beaker debug jobs (#1792) allenai/open-instruct
Short Beaker debug jobs now carry a 30-minute minimum runtime and run preemptible; the non-preemptible default was attaching an 8-hour `minRuntime` guarantee to jobs whose own `--timeout` was 15 to 30 minutes, and the author measured a 24+ minute queue on ai2/jupiter for an identical 1-GPU job. Anything you script around Beaker debug launches gets a different scheduling class.
[ci] fix: Scale Qwen3-8B PPO Ascend nightly CI to 16 NPUs (#8531) verl-project/verl
The remaining verl work is test infrastructure: the Qwen3-8B PPO Ascend nightly moves to 16 NPU runners to work around host OOM, the nightly Ascend workflow drops its PR trigger, and Ascend CI workflows now print meminfo for memory-pressure diagnosis, none of which changes how you run verl.
Action items
- → Set core.max_run_final = true in MILES configs that must export and complete after a wall-clock stop, since resumable... allenai/open-instruct [plan]
