$ the-wire · showcase
Tensordict compiles masked indexing under fullgraph; ExecuTorch links WebGPU into the Wasm build
By RepoJournal · Filed · About PyTorch · Composed from the cited sources · methodology
15 people shipped this
peterdsharpe
@peterdsharpe
3 cited
albanD
@albanD
1 cited
Roman Rädle
@raedle
1 cited
Ivan Zaitsev
@izaitsevfb
William Wen
@williamwen42
ethanwee1
@ethanwee1
Bob Ren
@bobrenjc93
Jacob Szwejbka
@JacobSzwejbka
Mark Inderhees
@mark-inderhees
Gasoonjia
@Gasoonjia
Digant Desai
@digantdesai
github-actions[bot]
@github-actions[bot]
liangel-02
@liangel-02
George Hong
@georgehong
Jerry Mannil
@jerrymannil
The day's work is mostly graph-capture plumbing and packaging: Tensordict closed a cluster of torch.compile gaps that broke plain tensor idioms on TensorDict, while ExecuTorch made browser-side WebGPU inference possible in an opt-in build.
[BugFix] Compile td[mask] and td[mask] = 0 with fullgraph, and write a number into a tensordict holding MetaData pytorch/tensordict
Boolean-mask selection and assignment, td[mask] and td[done] = 0, used to fail with fullgraph=True on a TensorDict while the same indexing compiled on a plain tensor, and without fullgraph it broke the graph. td[mask], td.masked_select(mask) and, on a lazy stack, td[mask] = number now compile in one graph, and an eager bug where td[index] = 1 raised an AttributeError on a tensordict holding Met...
[Compile] Compile permute and get with a default on a lazy stack, and in-place math on a TypedTensorDict over one pytorch/tensordict
td.permute(...) and td.get(("a", "x"), None) failed to compile on a lazy stack even though both compile on a dense TensorDict, and add_/mul_ on a TypedTensorDict over a lazy stack fail on the torch 2.16 nightly. All of them now compile on torch 2.13, 2.14.1 and the nightly with static and dynamic shapes.
Run WebGPU models in browser Wasm (#23938) (#23938) pytorch/executorch
The opt-in ExecuTorch Wasm build could not run WebGPU-delegated models in a browser. Linking the WebGPU backend in behind EXECUTORCH_BUILD_WASM_WEBGPU (off by default), adding a process-wide initWebGPU() binding, and removing two op limits that large vision models hit makes that path work.
Deprecate torch.utils.model_dump public APIs with FutureWarning (#201214) pytorch/pytorch
torch.utils.model_dump was soft-deprecated earlier, but only its CLI warned. Every function in its __all__ now emits a FutureWarning. The notes state this is not BC-breaking: all public APIs keep working.
[Performance] Make `pad` faster #2861 pytorch/tensordict
The long tail is performance: tensordict.pad is 15 to 20% faster and TorchRL's split_trajectories 12% faster, while stacking plain TensorDicts in torch.stack is 9 to 23% faster, which is the path EnvBase.rollout and LeanRL's PPO hit every rollout. ExecuTorch's FFT prim outputs now receive the contiguous strides their metadata already promised, the WebGPU backend gained a shared-memory direct ke...
