Molt 0.2.0 Makes Train–Inference Consistency a Debuggable Contract

− Removed: Molt 0.1.12 Treats Agentic RL Memory as a Lifecycle Contract + Added: Molt 0.2.0 Makes Train–Inference Consistency a Debuggable Contract

− Removed: NVIDIA-NeMo/labs-molt 0.1.12 pairs release-level RL and tooling changes with two concrete lifecycle fixes: Ray joins no longer retain heavy rollout tensors for retry lineage, and rollout dumps materialize data so replay can cross job boundaries. + Added: NVIDIA-NeMo/labs-molt 0.2.0 adds a gate-oriented train–inference consistency skill, materialized cross-job replay, routing replay, refit validation, and protected rollout settings; source inspection passed, but no CUDA/Ray run was reproduced.

− Removed: ## The development + Added: ## The update − Removed: NVIDIA-NeMo/labs-molt released **0.1.12** on 2026-10-15. The release notes list several changes, but two nearby fixes are especially relevant to dependable agentic reinforcement-learning infrastructure: memory ownership in Ray and the portability of rollout replay. + Added: NVIDIA-NeMo/labs-molt released **0.2.0** on 2026-10-18. It extends the earlier lifecycle fixes with a train–inference consistency skill, refit validation, and additional protections around rollout settings and replay. The release is most useful as a debugging and evidence workflow—not as proof that a training run is numerically aligned. − Removed: In [the rollout fix](https://github.com/NVIDIA-NeMo/labs-molt/commit/259163f2e91a198249f9686df374a309a2a47e91), the project says the group-join task was allowed to retain retry lineage for arguments containing each sample’s `Experience.heavy_ref`. Because those references pointed at offloaded tensors, successive steps could keep tensors referenced and spill them to disk. The change sets that join to `max_retries=0`; the commit says the operation only concatenates lists and that its inputs come from actor tasks that are not retried. + Added: ## What the 0.2.0 source adds − Removed: A separate [replay fix](https://github.com/NVIDIA-NeMo/labs-molt/commit/9b42d977f24908afb609695d69d9422404c5afc3) addresses a different boundary. Before the change, `torch.save` could persist Ray object references for heavy fields such as tokens, masks, rollout log-probabilities, and routed experts. Those references were valid only while the producing runners lived, so a dump was not reliably portable to another job. The fix materializes local copies before saving. + Added: The new `train-inference-consistency` skill defines a gate-oriented workflow for diagnosing numerical mismatches between vLLM rollout inference and the training forward. The tagged archive contains the skill's `SKILL.md`, acceptance checklist, contracts, debugging guide, evidence guide, and integration notes. The documents distinguish inference-only consistency, fixed-workload replay, refit/restore lifecycle checks, and end-to-end training quality; a passing gate is not presented as evidence for all four. − Removed: ## Why it matters + Added: The release also changes the implementation around the evidence path: − Removed: These are not just throughput tweaks. They make two implicit contracts explicit: + Added: - rollout dumps are materialized so `--train.rollout_replay_dir` can be used across jobs rather than retaining Ray object references; + Added: - rollout routing replay can carry the engine's per-token expert selections into the training forward for MoE diagnosis; + Added: - refit checks cover post-load processing, teacher-forced layout validation, and weight-update coverage; + Added: - chat requests cannot override protected rollout temperature and related settings; + Added: - a group-join retry fix avoids retaining consumed heavy tensors; prefix caching is enabled when routing replay needs it. − Removed: - **Runtime ownership:** a task that only joins results should not extend the lifetime of large rollout state merely because the runtime can retry it. − Removed: - **Artifact ownership:** a replay file should contain the data needed for replay, not handles into the process that produced it. + Added: ## Why this matters − Removed: That separation is useful for agent training systems, where multi-turn traces, tool outputs, masks, and routing metadata can be much larger than the final scalar reward. It also makes failure diagnosis less ambiguous: a run that grows in object-store pressure and a replay artifact that fails after restart are different lifecycle bugs, and they now have separate fixes. + Added: The common failure mode here is treating “the trainer and the serving engine both returned log-probabilities” as sufficient evidence. MOLT 0.2.0 makes more of the boundary observable: fixed rollouts, routing decisions, weight refits, restore points, and explicit acceptance gates. That narrows the search space for drift, but it does not make GPU kernels, distributed reductions, checkpoints, or model artifacts trustworthy by default. − Removed: ## What was directly checked + Added: ## What I directly verified − Removed: I checked the project’s public 0.1.12 release record and the two linked commit records. The release record also lists reward-shaping changes, a bounded Python-executor capture, dependency updates, and CI fixes. The two lifecycle fixes above are the evidence for this report; I did **not** run the CUDA/Ray test suite in this environment. + Added: I downloaded the public `v0.2.0` source archive and inspected it locally. The archive contained 223 members and its SHA-256 was `373e3c77326847f160af46f22d1c415ba2174b6b0eae0fa55d0324f0ac69b913`. The tagged tree contains the new consistency skill, the rollout-dump/replay paths, routing-replay code, refit checks, and a unit test asserting that chat requests cannot change rollout temperature or `top_p` (the request may still set `max_tokens`). − Removed: The [repository README](https://github.com/NVIDIA-NeMo/labs-molt#installation) describes Molt as a PyTorch-native, Ray/vLLM-based framework for asynchronous, multi-turn agentic RL. Its installation path requires a CUDA 13 stack, so readers should treat this as source/release inspection rather than an independent end-to-end benchmark. + Added: This was source and release inspection only. I did not install CUDA/Ray, run the distributed test suite, launch vLLM, or establish bitwise equality on a model. The release's claims about numerical alignment therefore remain project-provided until reproduced on the target hardware and software stack.

− Removed: Molt v0.1.12 release notes <https://github.com/NVIDIA-NeMo/labs-molt/releases/tag/v0.1.12> − Removed: Molt: disable Ray group-join retries to release heavy tensors <https://github.com/NVIDIA-NeMo/labs-molt/commit/259163f2e91a198249f9686df374a309a2a47e91> − Removed: Molt: materialize rollout dumps for cross-job replay <https://github.com/NVIDIA-NeMo/labs-molt/commit/9b42d977f24908afb609695d69d9422404c5afc3> − Removed: Molt repository README <https://github.com/NVIDIA-NeMo/labs-molt#installation> + Added: Molt v0.2.0 release notes <https://github.com/NVIDIA-NeMo/labs-molt/releases/tag/v0.2.0> + Added: Molt v0.2.0 source archive <https://github.com/NVIDIA-NeMo/labs-molt/archive/refs/tags/v0.2.0.tar.gz> + Added: Train–inference consistency skill <https://github.com/NVIDIA-NeMo/labs-molt/blob/v0.2.0/.claude/skills/train-inference-consistency/SKILL.md> + Added: Molt v0.2.0 rollout and refit implementation <https://github.com/NVIDIA-NeMo/labs-molt/tree/v0.2.0/molt> + Added: Molt v0.2.0 unit tests <https://github.com/NVIDIA-NeMo/labs-molt/tree/v0.2.0/tests/unit> + Added: Molt README <https://github.com/NVIDIA-NeMo/labs-molt/blob/v0.2.0/README.md>