Who contributed to PyTorch on August 24, 2026?

7 developers shipped this update, including aswanth-07, khazic, PlatformLC, fxlin0261, twthighd, shoumikhin, and qti-chenweng.

What were the notable PyTorch updates?

[FSDP2] Guard _unsharded_param access in to_accumulated_grad_if_needed (#194058), Promote commits to viable/strict on a schedule, not by commit count (#22064), and [BugFix] Preserve TruncatedNormal sample dtype (#4137).

@pytorch

PyTorch and the broader machine-learning ecosystem

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

FSDP2 CRASH FIXED, EXECUTORCH BUILD PIPELINE REPAIRED AFTER COMMIT PROMOTION BUG

By RepoJournal · Filed 06:58 UTC on August 24, 2026 · About PyTorch

7 people shipped this

PlatformLC @PlatformLC 2 cited

aswanth-07 @aswanth-07 1 cited

khazic @khazic 1 cited

fxlin0261 @fxlin0261 1 cited

twthighd @twthighd 1 cited

shoumikhin @shoumikhin 1 cited

qti-chenweng @chenweng-quic

PyTorch's distributed training guard catches an AttributeError that was silently breaking gradient accumulation in sharded models, while ExecuTorch's nightly builds restart after a scheduling bug locked out half of all commits.

The FSDP2 fix [1] addresses an unguarded access to `_unsharded_param` in `to_accumulated_grad_if_needed` that threw AttributeError when the gradient was missing, a dangerous condition in distributed training. Separately, ExecuTorch's viable/strict promotion system [2] was advancing only from commits whose position in history was divisible by 4, leaving every other commit permanently ineligible for nightly builds and breaking the manual promotion escape hatch. The TorchRL dtype preservation fix [3] ensures `TruncatedNormal.rsample()` respects the location tensor's dtype instead of upcasting through the global default, correcting a silent behavior change that could break downstream sampling code. Across testing, the team hardened distributed test infrastructure: PrivateUse1 accelerators now get microbatch test coverage [4], backend availability checks use the unified `c10d.is_backend_available` API instead of hardcoded dicts [5], and test_store classifies all 17 test cases with hardware labels [6]. A CUDA multinomial optimization [7] removes dead code from the without-replacement path, eliminating an unnecessary device-to-device copy.

Action items

→ Deploy FSDP2 fix to prevent gradient accumulation crashes in distributed training pytorch/pytorch [immediate]
→ Land ExecuTorch viable/strict promotion fix to restore nightly build eligibility pytorch/executorch [immediate]
→ Pull TorchRL dtype preservation fix to prevent sampling behavior regression pytorch/rl [plan]

References

[1] [FSDP2] Guard _unsharded_param access in to_accumulated_grad_if_needed (#194058) pytorch/pytorch
[2] Promote commits to viable/strict on a schedule, not by commit count (#22064) pytorch/executorch
[3] [BugFix] Preserve TruncatedNormal sample dtype (#4137) pytorch/rl
[4] [Testcase Refactoring] Enable PrivateUse1 tests and add hw-classification in test_microbatch (#190422) pytorch/pytorch
[5] [Testcase Refactoring] Use is_backend_available for MultiProcContinuousTest backend check (#192186) pytorch/pytorch
[6] [Testcase Refactoring] Add hw_classification to test_store (#192036) pytorch/pytorch
[7] [CUDA] Remove unused multinomial probability copy (#194391) pytorch/pytorch

Quick answers

What shipped in PyTorch on August 24, 2026?: PyTorch's distributed training guard catches an AttributeError that was silently breaking gradient accumulation in sharded models, while ExecuTorch's nightly builds restart after a scheduling bug locked out half of all commits. In total, 42 commits and 3 pull requests landed.
Who contributed to PyTorch on August 24, 2026?: 7 developers shipped this update, including aswanth-07, khazic, PlatformLC, fxlin0261, twthighd, shoumikhin, and qti-chenweng.
What were the notable PyTorch updates?: [FSDP2] Guard _unsharded_param access in to_accumulated_grad_if_needed (#194058), Promote commits to viable/strict on a schedule, not by commit count (#22064), and [BugFix] Preserve TruncatedNormal sample dtype (#4137).

PyTorch's Caffe2 cleanup hits a wall as internal dependencies force three reverts

PyTorch's attempt to shed its Caffe2 legacy just got blocked by internal dependencies, and the resulting reverts ripple across the cleanup effort.

python 58 shipped

@pytorch Aug 22

Pallas HBM Windows Open on Any Tensor Axis

Pallas stagers just gained the ability to align HBM windows on any tensor axis, not just the last one, and a critical autotuning crash is fixed.

python 81 shipped

@pytorch Aug 20

NVFP4 GroupedExperts lands on Blackwell

TorchAO's new differentiable NVFP4 grouped GEMM stack brings high-throughput grouped experts to Blackwell, and it's the headline act.

+10

python 160 shipped

@pytorch Aug 19

Pytorch Fixes XPU Dispatch Gap and CI OOMs

PyTorch's XPU backend lost a critical dispatch for _safe_softmax, and CI is getting a GOMAXPROCS pin to stop GPU nodes from OOMKilling.

+10

python 85 shipped

Elsewhere on the wire

AI Agents Aug 24

CLINE SHIPS V4.1.15 AS TELEMETRY BUGS SURFACE ACROSS THE AGENT STACK

Cline's long-awaited auto-approve fix landed overnight, but it's the telemetry drop in 4.1.11 that should have you checking your session metrics.

ai-agents 44 shipped

Local LLMs Aug 24

VLLM'S P2P WEIGHT SYNC CUTS RL MOE TRANSFER COSTS

Forget all-gathering experts during RL rollout: vLLM's new sharded_rdt backend lets each worker pull only the slice it needs, turning a total_bytes transfer into a total_bytes divided by num_workers one.