RepoJournal
Alignment & Post-training Alignment & Post-training
47 wires and counting

$ follow Alignment & Post-training

Keep up with Alignment & Post-training in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-11-16
stories 8

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

verl drops the V0 trainer, V1 becomes the only PPO path

By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology

verl removed its V0 PPO trainer from mainline and moved the legacy fully async and one-step-off implementations into verl-recipe, so any training config pointed at V0 has to move with it.

[BREAKING][trainer, recipe, ci, doc] refactor: remove V0 trainer and migrate legacy async trainers to recipes verl-project/verl

by Begunner

V1 is now the only PPO trainer in verl: the V0 trainer is gone from the mainline, while the fully async and one-step-off implementations moved to verl-recipe, where the migrated recipes pin verl v0.9.1 at commit 1876b06. Shared code still used by mainline remains in verl, so a config that targeted V0 will not resolve against mainline after this.

[docker, ci, doc] feat: upgrade the ROCm image to vLLM v0.32.0 verl-project/verl

by PeterYang12

The ROCm image in docker/rocm/Dockerfile.rocm now builds on vLLM v0.32.0, matching the version verl's CUDA stack already uses, and the ROCm e2e workflow runs against verlai/verl:rocm7.14_torch2.12_release_1112. The workflow also sets HSA_TOOLS_DISABLE_REGISTER=1, because since vllm-project/vllm#56190 every vLLM worker registers a rocprofiler client and idle GPU processes busy-poll in the ROCm r...

[vllm] fix: support MTP weight sync with vllm ModelRunner V2 (#8681) verl-project/verl

by lxb007981

MTP weight sync now resolves both V1 drafter.model and V2 speculator.model, so weight loading, buffer updates and post-processing work under either vLLM ModelRunner, with the same fallback applied to quantized weight loading. If you run MTP with vLLM ModelRunner V2 in the ROCm or CUDA stack, this is the fix that makes weight sync land correctly.

Quick answers

What shipped in Alignment & Post-training on November 16, 2026?
verl removed its V0 PPO trainer from mainline and moved the legacy fully async and one-step-off implementations into verl-recipe, so any training config pointed at V0 has to move with it. In total, 4 commits and 4 pull requests landed.
Who contributed to Alignment & Post-training on November 16, 2026?
3 developers shipped this update, including Begunner, PeterYang12, and lxb007981.
What were the notable Alignment & Post-training updates?
[BREAKING][trainer, recipe, ci, doc] refactor: remove V0 trainer and migrate legacy async trainers to recipes, [docker, ci, doc] feat: upgrade the ROCm image to vLLM v0.32.0, and [vllm] fix: support MTP weight sync with vllm ModelRunner V2 (#8681).