$ the-wire · showcase
verl restructures rollout quantization; TPU paths gain runbooks in both repos
By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology
verl's QAT and low-precision rollout config moves from a single string to structured keys under a breaking flag, and both verl and open-instruct shipped TPU example and runbook tooling that lowers the hardware bar for GRPO smoke tests and MaxText SFT.
[BREAKING][megatron, vllm, sglang, rollout, doc] refactor: unify QAT and low-precision rollout verl-project/verl
The string `rollout.quantization=fp8` API is gone, replaced by structured `rollout.quantization.linear`, `moe`, `ignore`, and `config_file` settings, and Megatron QAT now covers `fp8`, `mxfp8`, `mxfp4`, `mxfp4_mxfp8`, and `nvfp4` with compressed-tensors-compatible trainerside exporters for FP8_BLOCK, MXFP8, MXFP4, and NVFP4. This is a breaking change: any recipe that sets the old quantization s...
[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 targets on-policy distillation, where the existing `forward_kl_topk` keeps only the teacher's top-k terms. The PR notes that truncated objective is not a KL, goes negative whenever Q > P, and is clamped with `clamp_min(0)`, which silently drops the gradient; the new mode adds a tail term instead. Available across the fsdp, megatron, and veomni backends.
[hardware, doc] feat: TPU examples: public-package image, 2-chip GRPO smoke test, DP shard size from TP, SFT torchrun fixes verl-project/verl
The TPU examples now build the image only from public packages, and the Qwen3-0.6B GRPO smoke test needs 2 chips (2x `v6e-1`) instead of 8 (2x `v6e-4`), one trainer and one rollout chip, with `ray-tpu-v6e4-2slice.yaml` renamed to `ray-tpu-v6e-2slice.yaml`. Also folded in: DP shard size derived from TP size and SFT torchrun fixes. Only `examples/tpu/` changes.
TPU SFT runbook and tpu skill allenai/open-instruct
A runbook plus `.claude/skills/tpu` for Olmo 3/3.5 SFT on TPU via MaxText and Tunix, covering cluster access, queues, pinned images, checkpoint conversion, data prep, launch and monitoring, loss interpretation, resume, HF export, and known mesh constraints. Merge the companion scripts PR before or alongside it, since the guide cites them.
Add TPU MaxText SFT tooling with portable defaults allenai/open-instruct
The scripts behind that runbook: pinned image patches, Kubernetes overlays, dataset preparation, tokenizer parity checks, checkpoint conversion, and HF export. Dataset builders resolve a single HF revision or explicit local parquet inputs, honor standard caches, and refuse nonempty shard directories, which prevents accidental appends to an existing run's shards.
Action items
