Who contributed to PyTorch on August 17, 2026?

6 developers shipped this update, including gtnv, unography, digantdesai, Erik Lundell, Nikita Shulga, and Isalia20.

What were the notable PyTorch updates?

[BE] Removing dead code from MPS (5/5) (#193734), [BE] Removing dead code from MPS (4/5) (#193733), and [BE] Removing dead code from MPS (3/5) (#193732).

@pytorch

PyTorch and the broader machine-learning ecosystem

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

PYTORCH CUTS MPS DEAD CODE, DREAMER V3 FIXES GPU MEMORY LEAK

By RepoJournal · Filed 07:07 UTC on August 17, 2026 · About PyTorch

6 people shipped this

Isalia20 @Isalia20 4 cited

gtnv @gtnv 2 cited

unography @unography 2 cited

digantdesai @digantdesai 2 cited

Erik Lundell @Erik-Lundell 2 cited

Nikita Shulga @malfet 1 cited

PyTorch is aggressively pruning unused Metal Performance Shaders infrastructure while TorchRL patches a critical memory leak that kept loss tensors pinned to CUDA for entire training runs.

The PyTorch team merged a five-part series stripping dead code from MPS, removing unused shader helpers, kernel dispatch machinery, and graph cache formatting utilities that lost their callers as operations migrated to TensorIterator [1] [2] [3] [4]. In parallel, they consolidated MPS test coverage by running MPS-only tests in normal CI pipelines instead of gating them behind macOS-specific shards, reducing redundant test infrastructure [5]. Over on TorchRL, a critical bugfix eliminates GPU memory retention by ensuring loss tensors don't persist in CUDA memory across the full training duration and allowing Dreamer v3 to use available accelerators by default [6] [7]. The RL team also landed comprehensive numerical contract tests for TQC covering pooled target truncation, quantile-Huber loss gradients, and terminal-NaN masking [8], plus a performance optimization skipping unnecessary reshapes in MuJoCo playground's 1-dimensional batch inference path [9]. ExecuTorch completed the Muse Glimmer rename across test documentation and comments [10] [11], and the Arm backend now prefixes BSS sections to reduce object file bloat by letting the compiler optimize zero-initialized data at compile time rather than linker time [12] [13].

Action items

→ Upgrade TorchRL to latest and rebuild Dreamer v3 training jobs to recover CUDA memory pytorch/rl [immediate]
→ Review MPS test consolidation changes if you run macOS CI pipelines pytorch/pytorch [plan]
→ Monitor Arm ExecuTorch builds for BSS section improvements in binary size pytorch/executorch [monitor]

References

[1] [BE] Removing dead code from MPS (5/5) (#193734) pytorch/pytorch
[2] [BE] Removing dead code from MPS (4/5) (#193733) pytorch/pytorch
[3] [BE] Removing dead code from MPS (3/5) (#193732) pytorch/pytorch
[4] [BE] Removing dead code from MPS (2/5) (#193731) pytorch/pytorch
[5] Run MPS-only tests in normal test selection (#193586) pytorch/pytorch
[6] [BugFix] Fix dreamerv3 GPU usage, loss tensors in memory ↗ pytorch/rl
[7] [BugFix] Fix dreamerv3 GPU usage, loss tensors in memory (#4118) pytorch/rl
[8] [Test] [TQC] [3/3] Add TQC numerical contracts test ↗ pytorch/rl
[9] [Performance] mujoco playground skip 1d reshape ↗ pytorch/rl
[10] [ExecuTorch] Complete Muse Glimmer rename ↗ pytorch/executorch
[11] [ExecuTorch] Complete Muse Glimmer rename (#21863) pytorch/executorch
[12] Arm backend: Add .bss prefix to bss sections (#21839) pytorch/executorch
[13] Arm backend: Add .bss prefix to bss sections ↗ pytorch/executorch

Quick answers

What shipped in PyTorch on August 17, 2026?: PyTorch is aggressively pruning unused Metal Performance Shaders infrastructure while TorchRL patches a critical memory leak that kept loss tensors pinned to CUDA for entire training runs. In total, 20 commits and 5 pull requests landed.
Who contributed to PyTorch on August 17, 2026?: 6 developers shipped this update, including gtnv, unography, digantdesai, Erik Lundell, Nikita Shulga, and Isalia20.
What were the notable PyTorch updates?: [BE] Removing dead code from MPS (5/5) (#193734), [BE] Removing dead code from MPS (4/5) (#193733), and [BE] Removing dead code from MPS (3/5) (#193732).

PYTORCH SLASHES B200 BENCH TIME 5X

The B200 microbenchmark suite just went from a 7.5-hour serial slog to a 143-minute parallel run, and that's only the start of what's moving across PyTorch today.

python 92 shipped

@pytorch Aug 16

PYTORCH CORE FIXES CUDA MEMORY LEAKS AND CHECKPOINT STORAGE SHARING

PyTorch shipped critical fixes for fake tensor leaks and checkpoint deserialization on meta devices, while the RL stack tackled sparse masked modes and routing tie-breaking across three repos.

python 31 shipped

@pytorch Aug 15

TENSORDICT 0.14 LANDS WITH BREAKING API DEFAULTS, TORCHTITAN GAINS ASYNC-TP BACKEND

TensorDict shipped 0.14.0 with backward-incompatible changes to `to_module()` and `copy_at_()` defaults, forcing TorchRL to patch every call site before its own 0.14 release.

+12

python 84 shipped

@pytorch Aug 10

DYNAMO TIGHTENS PYTHON SEMANTICS AS PYTORCH SHIPS TEST REFACTOR WAVE

PyTorch's compiler caught three attribute-handling bugs in a single commit batch, closing loopholes where graph breaks leaked through collections.deque writes and unbound comparison operations.