$ the-wire · showcase
verl adds batch-invariant actor forward, MXFP4 MoE recipe lands
By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology
verl shipped the training-side half of exact 0-diff train-rollout log-probs and an MXFP4 recipe for MoE routed experts, while open-instruct fixed a reward verifier that was zeroing correct answers.
[fsdp, trainer, cfg, doc, tests] feat: batch-invariant actor forward and sampler log-prob formula verl-project/verl
An opt-in FSDP engine option, `batch_invariant` (default off), installs vLLM's batch-invariant kernel overrides (matmul, softmax, mean, RMSNorm plus the cuBLAS and TF32 settings) and computes log-probs with the vLLM sampler formula: fp32 `log_softmax` over the full vocabulary, then gather. Turn it on if you want train and rollout log-probs to agree; leave it off and nothing about your current r...
Take the last declared multiple-choice answer and read row-requested formats allenai/open-instruct
`mcqa_final_choice` used to zero out correct answers that repeated the letter, wrapped it in bold, parentheses or `\text{}`, or glossed it on the same line, and could not read the per-row answer formats a prompt actually requests. It now takes the last explicit declaration in the final answer and pulls the row's one-capture regex from `template_metadata.output_regex`, so reward scores on prepar...
[megatron, doc] feat: MXFP4 MoE recipe and Qwen3-30B-A3B example verl-project/verl
`examples/mxfp4/run_qwen3_30b_a3b_megatron.sh` runs GRPO on Qwen3-30B-A3B with DAPO-Math-17k on a single 4-GPU Blackwell node, tested on 4×B300 (4×B200 is untested). It reuses the data, hyper-parameters and layout of the MXFP8 recipe: training TP2 × EP2, full recompute, optimizer on CPU.
[algo, megatron, cfg, doc, tests] feat: score centering on the Megatron actor verl-project/verl
Score centering, previously FSDP-side, now runs on the Megatron actor: sampler head log-probs are normalized over the vocab-parallel shards, with the row max and normalizer all-reduced over the TP group and the heads split by context parallel the same way top-k distillation does. If you run Megatron and were blocked from enabling score centering, that restriction is gone.
Prepare Olmo 3.5 release data fixes allenai/open-instruct
On the long tail: open-instruct added CPU preparation tooling for the Olmo 3.5 release-data fixes (24,906 audited exclusion IDs, 58 replacement identity seeds, a knowledge-cutoff check on the release tokenizer), and V1 sync PPO picked up a first proof-of-concept for custom dataloaders.