$ the-wire · showcase
DreamerV3 gains compile auto-selection, background replay, and native resume
By RepoJournal · Filed · About PyTorch
TorchRL's DreamerV3 path and AsyncBatchedCollector shipped user-facing training controls this week, while tensordict fixed a silent data-loss bug in concatenation.
TorchRL's DreamerV3 now auto-selects its fastest compilation strategy via `optimization.compile=auto`, which on CUDA picks a compiled full learner step, RSSM scan, and CUDA graph, while CPU defaults stay eager [1]. DreamerV3 can also save and resume learner updates, counters, and logging through checkpoint utilities and native replay serialization; saving remains opt-in and existing training defaults are unchanged [2]. New public configuration exposes discrete actions, configurable environments, and a `DreamerV3DiscreteActorConfig` [3]. Separately, `AsyncBatchedCollector.start()` now collects directly into replay on a dedicated writer thread, letting the caller train independently; ordinary iteration stays synchronous and yields `None` [4]. Hydra configs now expose `auto_register_policy_transforms`, `pre_collect_hook`, `post_collect_hook`, and `compact_obs` for both multi-collectors [5].
In tensordict, concatenating `NonTensorData` and `NonTensorStack` failed dispatch, and concatenating two `NonTensorData` inputs with different values silently repeated the first input's value [6]. That path now routes through value-preserving concatenation. This is a bug fix worth rolling in if you pad non-tensor batches with non-tensor capture enabled.
Elsewhere, torchtitan added a `qwen3_6` config registry reusing the Qwen3.5 implementation for Qwen3.6-27B and Qwen3.6-35B-A3B [7], and ExecuTorch moved its CI pull workflow to linux_job_v3, swapping EC2 runner labels for OSDC equivalents [8].
Action items
- → Pull in the tensordict concatenation fix before relying on non-tensor batch padding pytorch/tensordict [immediate]
- → Test DreamerV3 with optimization.compile=auto on CUDA to pick up the faster default pytorch/rl [plan]
- → Adopt native replay checkpointing in DreamerV3 training if you currently resume without replay state pytorch/rl [plan]
- → Monitor the ExecuTorch CI runner migration for any label or job failures pytorch/executorch [monitor]
References
- [1] [Feature] Pick the fastest DreamerV3 compile strategy by default ↗ pytorch/rl
- [2] [Feature] Resume DreamerV3 with native replay checkpoints ↗ pytorch/rl
- [3] [Feature] Support configurable DreamerV3 environments and run controls ↗ pytorch/rl
- [4] [Feature] Support background replay collection in AsyncBatchedCollector ↗ pytorch/rl
- [5] [BugFix] Expose multi-collector options in Hydra configs ↗ pytorch/rl
- [6] [BugFix] Preserve non-tensor values during concatenation ↗ pytorch/tensordict
- [7] [models] Add Qwen3.6 config registry ↗ pytorch/torchtitan
- [8] Move pull.yml to linux_job_v3 (#22247) ↗ pytorch/executorch
