Who contributed to Alignment & Post-training on October 17, 2026?

3 developers shipped this update, including zpltys, aoshen02, and farhatkevin.

What were the notable Alignment & Post-training updates?

[rollout, vllm] fix: reject engine_kwargs that override verl-managed vLLM args (#8260), Add experimental MILES on-policy distillation with a Megatron learner, and [trainer] fix: use 'enable_switch' instead of 'hybrid_engine' to decide whether create rollout instance on hybrid resource.

Oct 17, 2026

$ the-wire · showcase

verl rejects conflicting vLLM engine_kwargs, open-instruct adds MILES OPD behind a Megatron learner

By RepoJournal · Filed 07:00 UTC on October 17, 2026 · About Alignment & Post-training · Composed from the cited sources · methodology

3 people shipped this

farhatkevin @farhatkevin 2 cited

zpltys @zpltys 1 cited

aoshen02 @aoshen02 1 cited

Two operator-facing changes land today: verl closes a silent config conflict in rollout, and open-instruct makes on-policy distillation selectable with training.algorithm = "opd".

[rollout, vllm] fix: reject engine_kwargs that override verl-managed vLLM args (#8260) verl-project/verl

by aoshen02

verl now rejects vLLM args passed through rollout.engine_kwargs.vllm that duplicate values it derives itself (TP, DP, expert parallelism, LoRA, routing replay). Previously the winner depended on where verl wrote the derived value: tensor_parallel_size from engine_kwargs replaced verl's, while enable_expert_parallel was always overwritten with expert_parallel_size > 1, so an explicit engine_kwar...

Add experimental MILES on-policy distillation with a Megatron learner allenai/open-instruct

by farhatkevin

The MILES integration gains an experimental on-policy distillation path using a Megatron learner and a managed SGLang teacher, selected by training.algorithm = "opd". Qwen3.5 needs the Megatron backend because OLMo-core supports only llama, qwen3, olmo2 and olmo3, so existing default-backend runs are unaffected. It replaces the older robertb/miles-qwen35-opd and claude/miles-* stack, which the ...

[trainer] fix: use 'enable_switch' instead of 'hybrid_engine' to decide whether create rollout instance on hybrid resource verl-project/verl

by zpltys

The trainer now consults enable_switch rather than hybrid_engine when deciding whether to create a rollout instance on hybrid resource, so hybrid deployments should verify which signal their config actually sets. Operators running hybrid resource setups are the ones who will see a behavior difference.

Build the MILES runtime image for linux/amd64 explicitly (#1973) allenai/open-instruct

by Kevin Farhat

The MILES runtime image is now built for linux/amd64 explicitly; on an arm64 Mac, a docker build without --platform resolved the local olmo-miles base tag for the host, found no arm64 variant and fell back to pulling docker.io/library/olmo-miles, which fails. build_image_and_launch.sh already passes --platform=linux/amd64, so the fix matters if you invoke docker build directly.