$ the-wire · showcase
verl defaults Ascend RL mode, open-instruct unblocks SFT packing
By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology
Two operator-facing fixes landed: verl enables vLLM-Ascend RL mode by default on NPU, and open-instruct stops rejecting valid packed datasets whose rows end without an EOS token.
Skip the extra-dir trailing-EOS check under metadata boundaries allenai/open-instruct
When preparing vLLM engine arguments on Ascend, verl now defaults rl_config.enabled to true, which disables NZ weights for sleep/wake and weight updates across actor and reward-model rollouts; explicit user overrides are preserved, so check any config where you set it false deliberately. vLLM-Ascend 0.30 also makes several NPU patch shims unnecessary, since 0.30's DeepseekV4MoE builds the hash-...
Skip the extra-dir trailing-EOS check under metadata boundaries (#2211) allenai/open-instruct
Under --document_boundaries_from_metadata, the default, the packer reads row boundaries from token_ids_part_*.csv.gz, so no file tail is dropped and _check_extra_numpy_dir's trailing-EOS test only rejected valid data. It had stopped a C14 SFT run at startup because olmo35-template tool-call rows end in <|im_end|>\n with no <|endoftext|>; if your packed data hits that failure, this is the fix, a...
Add named pools with independent sampling policies and prompt replay (#2187) allenai/open-instruct
RL runs get named pools with independently configured sampling and optional prompt replay, plus per-pool reporting of attempts, coverage and committed training exposure so you can see which prompt groups actually reach training. Named pools define membership and do not imply a balancing objective; the default remains one finite global shuffle, with uniform, weighted and adaptive exposure sampli...
[trainer] fix: exclude all-padding rows from per-sequence off-policy metrics (#7785) verl-project/verl
Per-sequence off-policy metrics were averaged over all rows, so each all-padding row left by _balance_batch contributed ppl=1, log_ppl_diff=0 and rho²=1, drifting your rollout-correction numbers further the more padding the batch needed. Padding rows are now skipped; token-level metrics are unchanged.
Action items
