$ cat alignment/week/2026-10-19.log
the week in review · Oct 19 – Oct 25, 2026
verl breaks out unified QAT and low-precision rollout
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
The refactor adds trainer-side exporters for FP8_BLOCK, MXFP8, MXFP4 and NVFP4, and SGLang refit grows MXFP8, MXFP4 and NVFP4 weight paths.
[BREAKING][megatron, vllm, sglang, rollout, doc] refactor: unify QAT and low-precision rollout verl-project/verl
Megatron QAT and low-precision rollout now sit on one trainer-side weight quantization and rollout-side refit boundary, with compressed-tensors-compatible exporters for FP8_BLOCK, MXFP8, MXFP4 and NVFP4 plus ModelOpt-backed Megatron QAT. This is the config-breaking item of the week: check your existing quantization settings against the new boundary before you upgrade.
[algo, fsdp, megatron, veomni] feat: add tail-aware top-k KL for on-policy distillation verl-project/verl
A new forward_kl_topk_tail loss mode keeps the mass the truncated forward_kl_topk objective drops when the teacher's top-k probability sums to less than one. If you run on-policy distillation and have been carrying the truncation bias, this is a loss change you can select now across the algo, FSDP, Megatron and VeOmni paths.
[2/3][sglang] feat: support mxfp8/mxfp4/nvfp4 weight refit verl-project/verl
SGLang rollouts can now refit MXFP8, routed-expert MXFP4 and routed-expert NVFP4 weights online: the trainer exports pre-quantized checkpoint tensors and a bucket-aware loader stages them. Pair it with the DeepGEMM kernel standardization to keep kernel and checkpoint formats consistent across vLLM and SGLang.
[3/3][vllm] feat: standardize low-precision kernels on DeepGEMM verl-project/verl
Supported vLLM FP8, MXFP8 and MXFP4 Linear and MoE rollout kernels now standardize on DeepGEMM, matching the SGLang backend policy. If you pin kernel behavior or tuned around the old paths, verify your kernel selection after updating.
[veomni,ci] feat: add uv-based veomni training backend and cover it in GPU CI (#8274) verl-project/verl
VeOmni is now a first-class uv training backend on the uv-managed CUDA 13 image stack, with GPU CI covering it and a fix for the open image build behind a filtering egress proxy. Operators running VeOmni can move onto the standard backend path instead of the ad hoc install it needed before.
[megatron, trainer] fix: recipe-agnostic THD padding and FP8 padding of top-k distillation tensors (#8325) verl-project/verl
Recipe-agnostic THD padding and FP8 padding of top-k distillation tensors are fixed, which matters if you run top-k distillation under Megatron with context parallelism. Sequence and float8 packing are a quiet way to silently shift distillation loss gradients, so re-run a short comparison after updating.
Update Miles integration and runtime to v0.1.2 allenai/open-instruct
Open-Instruct's runtime moves to Miles v0.1.2, adapting the driver to the new training/publication and manager interfaces, initializing Ray before W&B and tracking before worker arguments are serialized. Operators need matching Miles and Open-Instruct versions; the refreshed native parser snapshot changes what parses.
TPU SFT runbook and tpu skill allenai/open-instruct
Alongside: a TPU SFT runbook and repo skill for Olmo 3/3.5 with MaxText and Tunix, MaxText SFT tooling with pinned images and tokenizer parity checks, an opt-in you_search backend, documented RL Zero chat templates, Ascend CI moved to a cluster-local PyPI cache, an HDO optimizer-offload fix, and a TPU GRPO image fix for the missing orjson import and tensordict pin.
$ ls alignment/week/ # the briefings behind this review
Keep up with Alignment & Post-training in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.
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.