$ the-wire · showcase
verl now fails loudly on overlong prompts instead of silently truncating them
By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology
verl's agent loop finally honors the data.truncation policy you set, so the default error mode raises instead of dropping the head of an overlong prompt, and the MILES Tiny examples moved to batch 2048 at LR 6e-6 with token-mean loss.
fix(agent_loop): honor data.truncation policy in _cap_text_prompt_length verl-project/verl
AgentLoopBase._cap_text_prompt_length used to left-truncate any prompt over the cap, keeping only the suffix regardless of the policy; it now reads data.truncation, so with error (the default) an overlong prompt raises ValueError rather than quietly discarding its leading instructions. If your agent runs depend on that silent clipping, they will start failing at the prompt cap.
[trainer] fix: conditionally fetch token-level rewards (#8565) verl-project/verl
PPO metric collection now only fetches token_level_rewards when they are actually produced, which removes missing-field errors with algorithm.use_kl_in_reward=false. KL-in-reward workflows keep their existing data-writing behavior.
MILES Tiny recipe: batch 2048, LR 6e-6 and token-mean loss allenai/open-instruct
Both maintained Tiny templates (tiny/node.toml and tiny/three-node.toml) now run 2,048 responses per optimizer update, 256 prompts by 8 responses, with constant LR 6e-6 and no warmup, and calculate_per_token_loss=true across all four maintained MILES examples with group-standard-deviation normalization still off. Baseline, router and throughput test expectations moved with them, so any numbers ...
[ci, vllm] test: add top-k sampling support checks to Ascend CI (#8706) verl-project/verl
The vLLM sampling test suite now runs in the Ascend CI workflow, asserting generated completions are nonempty and carry one logprob record per token.
[fsdp, perf] chore: deepseek v4 flash performance optimization solution for NPU (#8698) verl-project/verl
Elsewhere on the verl side: a DeepSeek V4 flash policy change for NPU training, with the redundant path controlled by use_no_sync_for_gradient_accumulation removed, plus delta sync landed for NPU checkpointing.
Action items