Who contributed to PyTorch on July 21, 2026?

9 developers shipped this update, including MOHAMMED HANAN M T P, cyy, JaimeLucero, Dmitry Nikolaev, PyTorch MergeBot, SaoirseARM, Jacob Stevens, and Scott Roy, and 1 more.

What were the notable PyTorch updates?

Revert "[Dynamo] Use getattr_static when probing internal attributes (#189524)", Remove dead build options and unused CMake variables (#190450), and CUDAGraph releases memory after capture error (#190348).

@pytorch

PyTorch and the broader machine-learning ecosystem

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

DYNAMO REVERTS BREAKING CHANGE, EXECUTORCH CLEANS UP DEPRECATED TYPES

By RepoJournal · Filed 07:00 UTC on July 21, 2026 · About PyTorch

9 people shipped this

cyy @cyyever 3 cited

Dmitry Nikolaev @dnikolaev-amd 1 cited

PyTorch MergeBot @pytorchmergebot 1 cited

SaoirseARM @SaoirseARM 1 cited

Jacob Stevens @JakeStevens 1 cited

Scott Roy @metascroy 1 cited

Warren Deng @warrendeng 1 cited

MOHAMMED HANAN M T P @hanu-14

JaimeLucero @JaimeLucero

PyTorch reverted a Dynamo optimization that broke internal tests, while ExecutorTorch is aggressively deprecating c10 shims in favor of standard library types.

In brief

Dynamo reverted getattr_static for probing internal attributes. It broke stuff. Sometimes the simplest fix is the right one.
ExecuTorch nuked c10 shim includes and redundant std type aliases. Then reverted it two commits later. Classic "wait, we still need that" energy.
CUDAGraph now releases memory after a capture error instead of hanging onto it. Less OOM surprises in error paths.
ExecuTorch codegen now emits std::string_view and std::optional directly. Modern C++ in, legacy wrappers out.
PyTorch ripped out dead build options and unused CMake variables. Every cleanup like this saves someone 20 minutes of "why does this flag exist" debugging.

The Dynamo team rolled back a getattr_static optimization [1] that was intended to reduce overhead but instead broke nested graph break tests. This is a tactical retreat, not a design failure, but it signals that the compile path still has rough edges at scale. Meanwhile, ExecutorTorch is making the hard push to emit std::string_view and std::optional directly [4] instead of torch::executor aliases, cleaning up deprecated type layers. That cleanup [5] caused internal build failures [6], which they immediately reverted, so expect a more surgical second pass. On the memory front, CUDAGraph now properly releases private mempools after capture failures [3], fixing a leak that could exhaust reserved segments over time. The pytorch/pytorch mainline also removed five dead CMake build options [2] with no remaining consumers, including USE_NNAPI, USE_PROF, and USE_TENSORRT. ExecutorTorch also removed the _use_edge_ops=False default [8], and the Arm backend cleaned up stale xfails [9]. Finally, benchmark caught and fixed a ROCm-specific random fallback that was regressing A100 accuracy [7].

Action items

→ Review your torch.compile() usage if wrapping bound methods - the revert [ref:5] restores old behavior pytorch/pytorch [monitor]
→ If you consume ExecutorTorch wheels, audit your c10 shim includes before the next cleanup lands [ref:9] pytorch/executorch [plan]
→ Upgrade to CUDAGraph memory fix [ref:4] if running capture-heavy workloads with error handling pytorch/pytorch [plan]
→ Verify benchmark suite accuracy on A100 with the DistillGPT2 fix [ref:11] pytorch/benchmark [monitor]

References

[1] Revert "[Dynamo] Use getattr_static when probing internal attributes (#189524)" pytorch/pytorch
[2] Remove dead build options and unused CMake variables (#190450) pytorch/pytorch
[3] CUDAGraph releases memory after capture error (#190348) pytorch/pytorch
[4] Emit std::string_view/std::optional from codegen (#21000) pytorch/executorch
[5] Remove c10 shim includes and redundant std type aliases (#20900) pytorch/executorch
[6] Revert "Remove c10 shim includes and redundant std type aliases" (#21050) pytorch/executorch
[7] Guard DistillGPT2 fallback_random override to ROCm only (#190306) pytorch/benchmark
[8] Remove _use_edge_ops=False (#20909) pytorch/executorch
[9] Arm backend: Remove stale rshift scalar xfail (#21037) pytorch/executorch

Quick answers

What shipped in PyTorch on July 21, 2026?: PyTorch reverted a Dynamo optimization that broke internal tests, while ExecutorTorch is aggressively deprecating c10 shims in favor of standard library types. In total, 71 commits and 20 pull requests landed.
Who contributed to PyTorch on July 21, 2026?: 9 developers shipped this update, including MOHAMMED HANAN M T P, cyy, JaimeLucero, Dmitry Nikolaev, PyTorch MergeBot, SaoirseARM, Jacob Stevens, and Scott Roy, and 1 more.
What were the notable PyTorch updates?: Revert "[Dynamo] Use getattr_static when probing internal attributes (#189524)", Remove dead build options and unused CMake variables (#190450), and CUDAGraph releases memory after capture error (#190348).

PYTORCH SHIPS BUILD FIX WHILE HELION TUNES H100 KERNELS TO DEFAULT

A critical build regression in cusparselt.cpp is now patched, while the kernel autotuner promotes its pointwise seed heuristic to production defaults on H100 and B200.

python 36 shipped 1-min read

@pytorch 2 days ago

HELION BACKEND REFACTOR UNLOCKS MULTI-DEVICE COMPILATION; DYNAMO FIXES FORWARD-AD TANGENT BUG

Helion shipped a four-part backend architecture overhaul that finally separates printer, lowering, and backend registration logic by device, eliminating the need to edit 17 core files for every new backend.

python 51 shipped 2-min read

@pytorch 3 days ago

HELION SHIPS B200 AUTOTUNER, PYTORCH PLUGS CUDA DUPLICATION, TORCHRL ADDS PULL-BASED MONITORING

Helion's stacked reduction-seed redesign targets NVIDIA's B200 with dedicated SM100 heuristics, while PyTorch core eliminates duplicate CUDA wrappers in NCCL2 and TorchRL lands its first operational monitoring layer.

python 82 shipped 1-min read

@pytorch 5 days ago

EXECUTORCH SLASHES CONV LATENCY, PYTORCH CORE FIXES FLEXATTENTION SPECIALIZATION BUG

ExecutorTorch's ARM backend just folded scalar multiplication into convolution weights, eliminating post-conv ops entirely, while PyTorch core patches a FlexAttention bug that was over-specializing dynamic sequence lengths in block masks.

python 69 shipped 1-min read

Elsewhere on the wire

@cachyos about 9 hours ago

NORWAY MIRROR DOWN, KERNEL STACK SHIPS MAINTENANCE RELEASES

CachyOS infrastructure is reporting a downed Norway mirror while the kernel and toolchain stacks push routine version bumps across the board.

infra 55 shipped 1-min read

Elixir & Phoenix about 9 hours ago

ECTO GAINS FRAGMENT COLUMN NAMING, PHOENIX PLUGS GET SMARTER

Ecto's new with_columns/2 lets you assign names to SQL fragments, while Phoenix now handles module plugs with options in pipe_through declarations.

elixir 14 shipped 1-min read

Mobile platforms about 9 hours ago

KOTLIN RIPS OUT INTERPRETER GUTS, FLUTTER PATCHES RENDERING PIPELINE

Kotlin is systematically removing its IR interpreter in a multi-commit cleanup that signals a major shift in how the compiler handles intermediate representation, while Flutter ships critical Vulkan safety fixes and expands autofill support.

mobile 115 shipped 2-min read

@dotnet about 9 hours ago

MAUI 11 PREVIEW 6 SHIPS WITH 190 COMMITS, MACIOS FIXES GENERATOR FLAKINESS

.NET MAUI 11.0.0-preview.6 landed with significant platform improvements while dotnet/macios tackles a critical test parallelization bug that was silently dropping thread-safety calls from generated code.