$ cat pytorch/week/2026-11-09.log
the week in review · Nov 9 – Nov 15, 2026
torch.load weights_only path allows arbitrary file read
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
The security item outranks a heavy week: TorchRL 0.15 shipped, tensordict landed a batch of silent-wrong-answer fixes, and torch.load needs attention.
torch.load(weights_only=True) allows arbitrary file read pytorch/pytorch
The restricted unpickler is supposed to be the safe path for loading untrusted checkpoints, and it permits an arbitrary file read instead. Until a fix lands, any code that calls torch.load on weights you did not produce runs on an assumption the flag no longer backs up.
TorchRL 0.15: Dataset Exports, TD-MPC2, Robot Deployment, and Flow Q-Learning pytorch/rl
ReplayBuffer.export now writes out through Hugging Face and LeRobot adapters, joined by export_onnx_policy, MenagerieEnv and LowLevelController for robot deployment, plus flow Q-learning and one-step TD-MPC2. The release is flagged breaking, so read the upgrade notes before pinning.
[BugFix] Count the rows of a ragged lazy stack in numel() pytorch/tensordict
For a lazy stack with members of batch sizes [2], [3] and [1], the stack's batch size is (3, -1) and numel() returns 1 while reshape(-1) of the same stack has 6 rows. Anyone sizing buffers off numel() on ragged stacks was allocating against the wrong number.
[BugFix] clone() of a tensordict that has a device keeps -0.0 when its leaves are all floating pytorch/tensordict
clone() of a device-carrying TensorDict or tensorclass flipped every -0.0 of a floating leaf to 0.0 across float16, bfloat16, float32 and float64, on CPU and CUDA, eager and compiled. Reciprocal-based code silently changed sign-carrying infinities; it is quiet until one lands in a loss.
[BugFix] TensorDictParams computes with the loaded weights after load_state_dict(assign=True) pytorch/tensordict
A model keeping its parameters in TensorDictParams and reading them through torch.vmap kept computing with the old weights after load_state_dict(assign=True), provided one forward had already run, and nothing raised. parameters() showed the loaded values while the kernels used the old ones, which is the worst shape a checkpoint bug can take.
[perf] Move all-gather comm. on the FSDP compute stream (#5627) pytorch/torchtitan
TorchTitan moved all-gather communication onto the FSDP compute stream, a fix for CUDA allocation retries traced to memory fragmentation across streams that were visibly slowing training steps. If your FSDP runs show allocation retries at step boundaries, this is the knob family to look at.
[ROCm] Remove the ROCm clone_input override in CuFFTPlanCache (#199834) pytorch/pytorch
PyTorch had forced clone_input = true on ROCm since 2021 to work around hipFFT clobbering the input buffer of out-of-place transforms. Dropping it removes a full contiguous copy of the signal on every forward call, so ROCm FFT workloads stop paying that tax.
[BugFix] Let Ctrl-C, sys.exit() and generator.close() pass through a with-block on a tensordict pytorch/tensordict
Ctrl-C, sys.exit(), asyncio task cancellation or generator.close() inside a with block on a tensordict raised RuntimeError: Converting a tensordict to boolean value is not permitted, in place of the original exception. Interrupt handling and cleanup paths stopped being reliable inside those blocks. Alongside: a dropped consolidated tensordict and a locked one with listed nested keys both held t...
$ ls pytorch/week/ # the briefings behind this review
Keep up with PyTorch in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.
fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.
Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.