$ the-wire · showcase
Inductor gains an FMA-contraction kill switch; TorchRL can export TensorDict policies to ONNX
By RepoJournal · Filed · About PyTorch · Composed from the cited sources · methodology
12 people shipped this
PyTorch added a documented way to turn off FMA contraction in Inductor for bit-reproducible kernels, while a new tensor-specific, seeded Helion pointwise path and a TorchRL ONNX exporter land as durable, opt-in levers rather than default behavior.
[inductor] Add triton.enable_fp_fusion flag to turn off FMA contraction (#202218) (#202218) pytorch/pytorch
The `triton.enable_fp_fusion` config flag turns off FMA contraction, which makes the same Inductor kernel produce bit-identical output across processes instead of diverging with whichever autotuned config CachingAutotuner happened to keep.
[Versioning] Load TorchRL's own checkpoints with allow_pickle=True pytorch/rl
Starting in `tensordict` 0.16, `TensorDict.load_memmap` and `TensorDict.load` change the default from `allow_pickle=None` to `allow_pickle=False`, which broke TorchRL's own checkpoints because optimizer state dicts in `Trainer` (`CKPT_BACKEND=memmap` and non-tensor metadata are persisted via pickle; TorchRL now loads its own checkpoints with `allow_pickle=True`.
[Versioning] Load TorchRL's own checkpoints with allow_pickle=True (#4551) pytorch/rl
That change is now merged as a commit, so anyone pinning TorchRL ahead of the `tensordict` 0.16 upgrade should expect memmap Trainer checkpoints to fail without it.
[Feature] MicroDuckEnv: robot-conformant policy_observation pytorch/rl
`MicroDuckEnv` gained an opt-in `policy_observation=True` flag that emits the exact 61-D vector the MicroDuck robot runtime feeds its policies at 50 Hz, so you can train against the deployed observation layout instead of hand-slicing the 56-D `observation`.
[Feature] export_onnx_policy: export TensorDict policies to ONNX with torch.onnx.export pytorch/rl
`torchrl.modules.export_onnx_policy(policy, path, example, *, out_keys=None, opset_version=18)` exports a TensorDict policy directly via `torch.onnx.export` with no tensor-level wrapper, calling the policy through tensordict's `dispatch` so the graph takes and returns TensorDicts. It also landed as a commit.
[xpu][fix]Enable memory-aware partitioned scatter on XPU (#195533) pytorch/pytorch
Elsewhere, `_build_scatter_memory_state()` now checks `xpu` alongside `cuda`; on XPU-only builds it had returned `None` and the partitioned-scatter pass fell back to an unconstrained budget, ignoring `partitioned_scatter_non_model_floor_bytes`, and TorchTitan stopped a rank-0 `RequestDispatcher` failure from silently stalling the engine loop by wiring an `on_background_failure` callback into `R...
Action items
