Who contributed to PyTorch on July 29, 2026?

9 developers shipped this update, including cyy, Tristan Rice, jansel, qti-horodnic, haowhsu-quic, SS-JIA, JakeStevens, and github-actions[bot], and 1 more.

What were the notable PyTorch updates?

Migrate deprecated torch.norm calls to torch.linalg.* (#185097), [c10d][nccl2] Implement gather_single (#191067), and [c10d] Rename gather_into_tensor to gather_single (#191073).

@pytorch

PyTorch and the broader machine-learning ecosystem

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

COLLECTIVE OPS STANDARDIZATION LANDS ACROSS C10D BACKENDS

By RepoJournal · Filed 07:00 UTC on July 29, 2026 · About PyTorch

9 people shipped this

Tristan Rice @d4l3k 4 cited

qti-horodnic @qti-horodnic 2 cited

cyy @cyyever 1 cited

jansel @jansel 1 cited

JakeStevens @JakeStevens 1 cited

haowhsu-quic @haowhsu-quic

SS-JIA @SS-JIA

github-actions[bot] @github-actions[bot]

jeanschmidt @jeanschmidt

PyTorch's distributed communication layer just got a consistency overhaul: gather_single now works on nccl2, collective timing surfaces across all backends, and a new NaN check hook works everywhere.

The c10d team shipped five coordinated changes that unify distributed collective operations across in-tree nccl2, nccl-lazy, gloo, and custom backends [2] [3] [4] [5]. The headline move: gather_single, the single-output-tensor analog of all_gather_single, finally works on the in-tree nccl2 and nccl-lazy backends [2]. The same PR renames the old gather_into_tensor API to gather_single with a deprecated alias, matching the pattern already used for all_gather_single [3]. Collective timing and work sequence numbers, which stock nccl provided but in-tree backends silently no-opped, now surface uniformly through enableCollectivesTiming(), getDuration(), and getSequencenumber() [4]. A new backend-agnostic NaN check hook builds on ProcessGroup pre-hooks instead of native implementations, so the debug feature works for any backend, not just ProcessGroupNCCL [5]. In the main pytorch/pytorch repo, 22 deprecated torch.norm call sites got migrated to torch.linalg.vector_norm, cleaning up the API surface ahead of the eventual full deprecation [1]. Helion's flash attention optimizations continue to land in stacked PRs, tuning two-CTA kernels for aligned noncausal fp16 HD64 shapes from 32K through 128K [6]. ExecuTorch added QNN backend support for sort and pdist operations [7] [8], plus bf16 delegation for XNNPACK fully-connected layers so bfloat16 models can lower their linears instead of falling back to portable [9].

Action items

→ Review gather_single migration if you use in-tree nccl2 or nccl-lazy backends pytorch/pytorch [plan]
→ Update torch.norm call sites if you maintain downstream code pytorch/pytorch [monitor]
→ Enable collectives timing and sequence tracking for profiling and hang debugging pytorch/pytorch [plan]

References

[1] Migrate deprecated torch.norm calls to torch.linalg.* (#185097) pytorch/pytorch
[2] [c10d][nccl2] Implement gather_single (#191067) pytorch/pytorch
[3] [c10d] Rename gather_into_tensor to gather_single (#191073) pytorch/pytorch
[4] [c10d][nccl2] Add collective timing and work sequence numbers (#191065) pytorch/pytorch
[5] [c10d] Add a backend-agnostic NaN check hook (#191064) pytorch/pytorch
[6] [cutedsl] Optimize dense HD64 two-CTA flash attention ↗ pytorch/helion
[7] Qualcomm AI Engine Direct - Adding QNN backend support for the sort core ATen op ↗ pytorch/executorch
[8] Qualcomm AI Engine Direct - Adding QNN backend support for _pdist_forward core ATen op ↗ pytorch/executorch
[9] [xnnpack] Support bf16 delegation for fully-connected ↗ pytorch/executorch

Quick answers

What shipped in PyTorch on July 29, 2026?: PyTorch's distributed communication layer just got a consistency overhaul: gather_single now works on nccl2, collective timing surfaces across all backends, and a new NaN check hook works everywhere. In total, 92 commits, 42 pull requests, and 3 releases landed.
Who contributed to PyTorch on July 29, 2026?: 9 developers shipped this update, including cyy, Tristan Rice, jansel, qti-horodnic, haowhsu-quic, SS-JIA, JakeStevens, and github-actions[bot], and 1 more.
What were the notable PyTorch updates?: Migrate deprecated torch.norm calls to torch.linalg.* (#185097), [c10d][nccl2] Implement gather_single (#191067), and [c10d] Rename gather_into_tensor to gather_single (#191073).

EXECUTORCH BACKENDS PATCH METADATA BUG, PYTORCH FIXES TEST BLOAT

Qualcomm and Arm backends shipped critical fixes to restore partition metadata and tame test runtime bloat across GPU CI.

python 17 shipped

@pytorch Aug 2

HELION RUNTIME REORGANIZED, DYNAMO FIXES NESTED GRAPH BREAKS, EXECUTORCH LANDS MOE QUANTIZATION PIPELINE

AmesingFlank completed a three-PR refactor moving Cute, Pallas, and Triton backend code out of Helion's runtime root into dedicated subpackages, while PyTorch's Dynamo team landed critical fixes for hook dict reconstruction and TLS attribute misses under nested graph breaks.

python 42 shipped

@pytorch Jul 27

EXECUTORCH ROUTES REASONING SEPARATELY, PYTORCH MODERNIZES COMPILER STACK

ExecutorTorch's shared OpenAI adapter now exposes optional reasoning content as a first-class output channel, while PyTorch core moves to clang-21 and strips dead C++ macros.

python 20 shipped

@pytorch Jul 26

DYNAMO FIXES GRAPH-BREAKING SDPA PARAMS, PAT SVD LEAK BREAKS DOWNSTREAM TESTS

PyTorch shipped fixes for two critical compilation and module patching bugs overnight that were silently breaking downstream workflows.

python 19 shipped

Elsewhere on the wire

AI Agents Aug 6

LANGCHAIN HARDENS ANTHROPIC TOOL CALLING AND GIT CREDENTIALS AS OPENHANDS SHIPS 1.10.0

LangChain fixed a critical credential leak in evaluation runs while Anthropic tool schemas get proper fallback handling, as three major agent frameworks push releases across the ecosystem.

ai-agents 54 shipped

Local LLMs Aug 6

OLLAMA FIXES KV CACHE CORRUPTION, LLAMA.CPP HARDENS FILE SEARCH, VLLM KILLS MOE LEGACY CODE

Ollama patched a critical cache corruption bug that corrupts lazy snapshots during multi-request prefills, while llama.cpp hardened Windows junction traversal and vLLM removed 14 hardcoded MoE shards across quantization methods.