RepoJournal
Alignment & Post-training Alignment & Post-training
16 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 2026-10-19
stories 6
RSS Archive Reviews

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

VeOmni becomes a first-class uv training backend in verl

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

ETOgaosion lxb007981
2 people shipped this
ETOgaosion @ETOgaosion 1 cited lxb007981 @lxb007981 1 cited

verl promoted VeOmni to a standalone uv-managed training backend on the CUDA 13 stack, put it under GPU CI, and tightened the Ascend install path, leaving operators with one adoptable change and no required migration.

[veomni,ci] feat: add uv-based veomni training backend and cover it in GPU CI (#8274) verl-project/verl

by Gao Ziyuan

VeOmni is now installable as its own backend on the universal uv-managed CUDA 13 stack via `uv run --frozen --all-packages --extra vllm --extra veomni`, with the `veomni` package pinned dependency-free in `uv.lock` and no `fsdp` or `veomni-sft` extras required. The revisions are pinned too: VeOmni 0.1.12 and Megatron-LM `core_v0`. GPU CI now covers the backend, so a stack you build on it gets e...

[ci] chore: use Ascend cluster-local PyPI cache (#8222) verl-project/verl

by lxb007981

Ascend CI jobs now set `PIP_INDEX_URL` and `PIP_TRUSTED_HOST` to route through the cluster-local PyPI cache, and the explicit PyPI index override is gone from the nightly VeOmni installation. If you mirror or override the index in your own Ascend pipeline, this is the shape verl expects now: the environment variables carry the cache, not a per-install flag.

$ ls alignment/reviews/ # the step-back reads

the week in review FSDP2 CPU offload corrupted parameters silently
Full archive → Weekly reviews →

Quick answers

What shipped in Alignment & Post-training on October 19, 2026?
verl promoted VeOmni to a standalone uv-managed training backend on the CUDA 13 stack, put it under GPU CI, and tightened the Ascend install path, leaving operators with one adoptable change and no required migration. In total, 3 commits and 3 pull requests landed.
Who contributed to Alignment & Post-training on October 19, 2026?
2 developers shipped this update, including ETOgaosion and lxb007981.
What were the notable Alignment & Post-training updates?
[veomni,ci] feat: add uv-based veomni training backend and cover it in GPU CI (#8274) and [ci] chore: use Ascend cluster-local PyPI cache (#8222).
♥

$ log Alignment & Post-training --recent

verl standardizes low-precision rollout kernels on DeepGEMM 20 shipped TPU GRPO image fixed, Megatron optimizer offload leaves HDO state alone 24 shipped verl restructures rollout quantization; TPU paths gain runbooks in both repos 10 shipped verl auto-aligns score centering with verified sampler support 7 shipped

$ log --all # elsewhere on the wire

Agent Security Guardrails exposes hooks over HTTP, certifi floor raised in agent-scan AI Red Teaming PyRIT fences judge transcripts against role forgery

$ ls showcases/

Safety Evals AI Red Teaming Agent Security Interpretability

$ init your-desk

RepoJournal can write one of these for any repo you care about. Sign in with GitHub, pick your repos and sources, and your desk gets a wire like this one every morning.