labs-molt: Skill for Claude Code

Train–inference consistency

Goal: for every valid response token, the logprob the trainer recomputes under the same weights equals the logprob vLLM returned when it sampled the token — the same bytes in the original dtype, mismatch/bitwise_frac = 0, at every RL step. Then bring rollout throughput back to within the performance budget: the aligned arm loses at most 20 % of the BI=0 arm's rollout tokens/s.

Where the code is: main has the measurement side — the mismatch/* metrics, rollout dump and replay, the engine and trainer flags in molt-integration.md. The aligned kernels behind --train.batch_invariant exist only on branch jianh/batch-invariant, the reference implementation (vLLM 0.28; Qwen2/Qwen3 dense and Qwen3-MoE). On main the task is therefore a port along training-side.md, never a flag flip. Read the installed vLLM inside a worker (which vLLM am I on); when it differs from the reference implementation's pin, say so before porting anything. Engine facts that depend on the installed version sit in one table with their re-verify checks: facts bound to the pinned vLLM.

Method: the training forward calls vLLM's own kernels for every operator that produces a value, so inference defines the numbers. The backward reuses the native math of the training stack under a written gradient contract; it is not bitwise. Molt's take on the idea, inspired by the vLLM blog post Finish RL Training Faster: Bitwise-Consistent, High-Performance, Agent-Driven (verl + Megatron): the gates C0–C10 match it one-to-one so the two can be read together, and every rule here is written for FSDP2 + AutoModel + vLLM. Scope: consistency only — day-0 model support, per-model case records and cluster notes are not part of this skill.

Start by task

Read the sections of your route in order; the rest of the skill is lookup material.