$ cat alignment/week/2026-11-02.log
the week in review · Nov 2 – Nov 8, 2026
verl fixes FP8 KV cache scales after quantized refit
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
NVFP4 training and rollout landed end to end too, and MILES now resumes after a wall-clock stop instead of exporting and exiting.
[vllm, rollout] fix: reset uncalibrated KV cache scales after a quantized refit verl-project/verl
Quantized rollout under FP8_BLOCK with an fp8_e4m3 KV cache made vLLM generate garbage because uncalibrated attention q/k/v/prob scales survived the refit; the fix resets them to 1.0 after every quantized refit. If you enabled FP8_BLOCK linear/moe rollout alongside an FP8 KV cache, that combination was broken until this landed.
[BREAKING][megatron, vllm] feat: per-token activations for NVFP4 QAT verl-project/verl
The static per-tensor activation scale is gone, so each token is quantized with a scale from its own amax, matching the NVFP4 rollout scheme. Existing QAT configs that tuned the static scale need rework before you resume.
[megatron, vllm, env] feat: add native NVFP4 training and rollout with live refit verl-project/verl
Native W4A4 training and rollout run on NVIDIA Blackwell, demonstrated with Qwen3-30B-A3B-Base on DAPO using Megatron and Transformer Engine with vLLM, with live refit between the two. That gives you a full NVFP4 path rather than inference-only quantization.
Overlap prepared refresh batches with training allenai/open-instruct
Ready-data preparation now starts while the current batch trains; the change notes it targets the roughly 8.7s exposed acquisition stage in an earlier three-node profile. Learning batch, lag limit, reward policy and datasets stay unchanged.
MILES: resumable wall-clock stop by default (core.max_run_final opts out) allenai/open-instruct
A stop triggered by core.max_run_seconds now resumes where it left off instead of exporting and completing. Set core.max_run_final = true if your jobs rely on the old behavior.
[hardware, megatron, doc] feat: upgrade Ascend 950 to Megatron-Core 0.18 and add MXFP8 GRPO example verl-project/verl
The Ascend 950 environment moves to Megatron-Core 0.18 and MindSpeed core_r0.18.0, with a Qwen3-30B-A3B MXFP8 GRPO example added. Operators there are upgrading from 0.12, so expect recipe differences.
[fully_async, ckpt, rollout] feat: decouple checkpoint publish and rollout commit verl-project/verl
CheckpointEngineManager gains a decoupled publish mode that returns before rollout commits the weights; coupled strong synchronization remains the default and preserves the NCCL/NIXL behavior you have today.
[rollout] feat: check Continuous Token merges against the chat template verl-project/verl
Continuous Token merges are now validated against the chat template, covering the cases where Qwen3 continues past <|im_end|> and GLM-4.7's <|observation|> is both a stop token and content. Alongside: TileLang FP8 and FP4 QAT kernels for FP8_BLOCK, MXFP8 and MXFP4, corrected Dolci IF sources, a 20-minute allowance for managed safety judges plus better SciCode error detail, and an Ascend nightly...
$ 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.