PyTorch 2.14.0 ships, TorchRL 0.14 lands

By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology

♥

TensorDict 0.14.1 follows as a correctness patch, and SECURITY.md now extends the distributed trust model to checkpoints.

1894 commits 905 PRs merged 22 releases 29 briefings covered

all pytorch reviews →

← previous month pytorch wire →

PyTorch 2.14.0 Release pytorch/pytorch

by ethche

The stable release ships with documented backwards-incompatible changes and deprecations, so read the notes before bumping. It also becomes the version the release matrix advertises as stable.

[SECURITY.md] Clarify that checkpoints share the distributed-features trust model (#195600) pytorch/pytorch

by Nikita Shulga

SECURITY.md now states explicitly that torch.distributed.checkpoint, including its format-conversion utilities, carries the same trust model as the rest of the distributed features. If you load checkpoint files from outside your own training job, that is now a documented boundary rather than an assumption.

Update security guidelines in SECURITY.md (#195610) pytorch/pytorch

by albanD

The same pass clarifies how to treat malformed de-serialized objects: PyTorch models should not be loaded from untrusted sources. Together with the checkpoint paragraph, this is the period's clearest statement of where deserialization risk sits.

TorchRL 0.14: DreamerV3, Async at Scale, and Composable Replay pytorch/rl

by vmoens

A major release spanning model-based RL, distributed training and inference, replay-buffer composition, checkpointing, and VLA workflows, with several contracts tightened. Anyone pinned to the pre-0.14 API surface should read those contract changes before upgrading.

TensorDict v0.14.1 pytorch/tensordict

by vmoens

A patch release fixing module metadata, lazy and non-tensor indexing, and frozen tensorclass deserialization. Despite the release's own breaking label, the notes state there are no breaking changes, new features, or deprecations.

Reject re-running `import torch` after a failed import (#194654) pytorch/pytorch

by albanD

A failed import now stays failed instead of silently re-running on the next import, and reload(torch) is forbidden outright; the notes describe the old behavior as doing weird things on the first call and segfaulting from the second. Relevant to test harnesses and notebooks that retry imports after an error.

Give causal attention the mask PyTorch asks for (#22443) pytorch/executorch

by Anthony Shoumikhin

Attention with is_causal=True was returning wrong values with nothing raised whenever the query was shorter than the keys, with errors against eager measured between 2.7 and 3.4 for a 1-query, 16-key shape. Silent numerical errors in causal attention are the kind that hide inside a training run.

[ROCm][BE] Remove outdated version conditions (#193927) pytorch/pytorch

by Yuanyuan Chen

ROCm 7.0 is now the minimum supported version, so guards for 6.3, 6.4, 6.5, and 7.0 were always-true and have been dropped. If your build scripts still branch on those versions, they are branching on dead code.

$ ls pytorch/month/ # the briefings behind this review

Wed Sep 2 ExecuTorch and TensorRT harden CUDA edge cases Thu Sep 3 PyTorch 2.14.0 ships, import reload guarded, AI reviewer unblocks merges Fri Sep 4 Inductor fixes stale captured-tensor mutations in while_loop Sat Sep 5 executorch fixes dropped edge-manager passes; TorchTitan overhauls MxFP8 Sun Sep 6 Torchtitan trains MoE routers in fp32 with bf16x9 on Blackwell Mon Sep 7 MPS complex solves land, Arm round() matches torch, dynamo fixes repro fidelity Tue Sep 8 AsyncEnvPool shared-memory fixes, CUDA 12.6 nightly removal, and stale autotune Future collision Wed Sep 9 DreamerV3 gains compile auto-selection, background replay, and native resume Thu Sep 10 TorchRL flips async collector defaults, GreenLight adds shadow state Fri Sep 11 TorchRL 0.14 lands, ExecuTorch tunes Arm64 BF16 copies and Vulkan dispatch Sat Sep 12 TorchTitan reshapes gradient accumulation for CUDA graphs, TensorRT 11.3 lands Sun Sep 13 PyTorch trims CuTeDSL known-good versions, fixes lerp on MPS Mon Sep 14 Helion autotunes Metal, TorchRL stops cacheing constructor device Tue Sep 15 Dynamo stops deleting globals it never created, FP8 blockwise scale fixes land Wed Sep 16 TorchRL resume lands, TorchTitan fixes a silent gradient bug Thu Sep 17 ROCm test timeouts traced to duplicated runs, and distributed/test_store drops 15 minutes Fri Sep 18 CPU caching allocator teardown race fixed, torchtitan shares one training engine Sat Sep 19 Fake-tensor hardening, ROCm graph fixes, and an XNNPACK use-after-free Sun Sep 20 torchtitan fixes multi-node HF export, torch._native opens to non-aten namespaces Mon Sep 21 TorchTitan drops pure TP on routed experts, restacks fused FFN weights Tue Sep 22 NVFP4 and MXFP4 quantization land in TensorRT, GreenLight tightens its merge bar Wed Sep 23 Dynamo stops silently folding with runtime_checkable Protocols, ExecuTorch drops CUDA 12.6 wheels Thu Sep 24 torch.compiler.precompile becomes a module, callable API retired Fri Sep 25 TorchTitan unifies graph execution and makes grouped MoE projections the default Sat Sep 26 Green-context SM splitting, cat-kernel DCE, and a debug map that pinned every exported module Sun Sep 27 TorchTitan pipeline and CP sharding land, ExecuTorch quantization finally holds Mon Sep 28 Inductor cat merge fixed, AVX512 grid_sample returned zeros Tue Sep 29 Inductor stops reinplacing index_put, ExecuTorch fixes ConvTranspose BatchNorm fusion Wed Sep 30 Inductor's nextafter lowering, cuDNN RNN strides, and a memory-snapshot refcount fix

all pytorch reviews →

← previous month pytorch wire →