Who contributed to PyTorch on July 16, 2026?

11 developers shipped this update, including Oscar Andersson, metascroy, zhaoxul-qti, Conarnar, Stephen Jia, Michael Lazos, Srijan Upadhyay, and drisspg, and 3 more.

What were the notable PyTorch updates?

Arm backend: Fold scalar mul into convolution (#20838), Support FP16 GGUF export (#20968), and Qualcomm AI Engine Direct - Add QNN Windows MSVC Build to CI.

@pytorch

PyTorch and the broader machine-learning ecosystem

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

EXECUTORCH SLASHES CONV LATENCY, PYTORCH CORE FIXES FLEXATTENTION SPECIALIZATION BUG

By RepoJournal · Filed 07:00 UTC on July 16, 2026 · About PyTorch

11 people shipped this

Oscar Andersson @oscarandersson8218 1 cited

metascroy @metascroy 1 cited

zhaoxul-qti @zhaoxul-qti 1 cited

Michael Lazos @mlazos 1 cited

Srijan Upadhyay @CodersAcademy006 1 cited

drisspg @drisspg 1 cited

Akash Verma @akashveramd 1 cited

Jake Stevens @JakeStevens 1 cited

Conarnar @Conarnar

Stephen Jia @SS-JIA

jinsooihm @jinsooihm

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.

In brief

ExecutorCh folds scalar multiplication directly into convolution ops on Arm. Cuts latency on the thing your phone actually runs. Sounds small, changes the benchmarks.
PyTorch fixes FlexAttention's dynamic sequence length specialization. It was over-specializing block masks, tanking reuse. One of those bugs that compounds quietly.
ExecutorCh now exports FP16 GGUF weights. Your ONNX pipeline doesn't care, but inference apps that need smaller models just got a cleaner export path.
PyTorch inductor swaps its NVGEMM codegen to use cuTLASS operators directly. Less custom glue, more leverage off Nvidia's library. Consolidation always wins in the long run.
ROCm gets native cholesky_ex support via hipSOLVER. PyTorch drops the Magma dependency for these tests. AMD's math library story gets a little less fragmented.

The ARM backend optimization [1] removes redundant multiplication operations by baking scalar and channel-wise scales directly into convolution weights and bias, cutting latency for models that chain conv layers with scaling operations. This lands alongside ExecutorTorch's new FP16 GGUF export path [2], which threads activation dtype through the entire pipeline so models can run fp16 compute instead of forced bf16, a critical feature for edge devices where every bit of memory matters. PyTorch core is fixing FlexAttention's over-aggressive specialization [4] that was pinning dynamic sequence lengths too early, breaking the contract for auto-dynamic dimension detection and forcing unnecessary recompilations. On the infrastructure side, ExecutorTorch added Windows MSVC CI validation for the Qualcomm AI Engine backend [3], ensuring QNN builds stay green on Windows without manual checks. ROCm users get a win too: cholesky_ex now works without MAGMA via hipSOLVER's new xgeev support [6], shipping in ROCm 7.14+. The NVGEMM backend migrated off the deprecated cutlass_api package onto the official cutlass.operators API [5], unblocking future NVIDIA CUTLASS DSL updates.

Action items

→ Review ExecutorTorch ARM conv folding if you're shipping edge models with multiplication chains pytorch/executorch [plan]
→ Update to latest XNNPACK if you hit numerically sensitive test failures [ref:10] pytorch/pytorch [monitor]
→ Patch multilabel_margin_loss if you're using -1 padded targets [ref:7] pytorch/pytorch [immediate]
→ Test FlexAttention dynamic shapes after the specialization fix lands pytorch/pytorch [plan]

References

[1] Arm backend: Fold scalar mul into convolution (#20838) pytorch/executorch
[2] Support FP16 GGUF export (#20968) ↗ pytorch/executorch
[3] Qualcomm AI Engine Direct - Add QNN Windows MSVC Build to CI ↗ pytorch/executorch
[4] Avoid specializing FlexAttention dynamic sequence lengths in block mask checks (#189952) pytorch/pytorch
[5] [inductor][NVGEMM] Migrate to cutlass.operators (#189772) pytorch/pytorch
[6] [ROCm] Enable cholesky_ex via hipSOLVER xgeev (ROCm >= 7.14) & remove Magma dependency for autogenerated tests (#188720) pytorch/pytorch
[7] [xplat] (Re-re-land) Update XNNPACK to github revision f6486e3e1d (#189871) pytorch/pytorch
[8] Fix multilabel_margin_loss decomposition for -1 padded targets (#189552) pytorch/pytorch

Quick answers

What shipped in PyTorch on July 16, 2026?: 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. In total, 60 commits and 9 pull requests landed.
Who contributed to PyTorch on July 16, 2026?: 11 developers shipped this update, including Oscar Andersson, metascroy, zhaoxul-qti, Conarnar, Stephen Jia, Michael Lazos, Srijan Upadhyay, and drisspg, and 3 more.
What were the notable PyTorch updates?: Arm backend: Fold scalar mul into convolution (#20838), Support FP16 GGUF export (#20968), and Qualcomm AI Engine Direct - Add QNN Windows MSVC Build to CI.

INDUCTOR FIXES NESTED REDUCTION BLOCK POINTER BUG, TORCHRL UNIFIES CHECKPOINTING

Inductor's Triton codegen was silently corrupting fused broadcasts and reductions by losing track of block pointer advancement across outer loops, now patched [ref:3].

python 77 shipped 1-min read

@pytorch 3 days ago

PYTORCH FIXES CRITICAL MPS MATH BUG WHILE TAMING CI GHOSTS

PyTorch's Metal backend shipped an accurate erfc function that eliminates 100% relative error on upper tail computations, fixing a cascade of broken numerics in GELU and special functions.

python 14 shipped 1-min read

@pytorch 4 days ago

PYTORCH FIXES H100 FLASH ATTENTION CI BREAKAGE, HELION CONSOLIDATES PALLAS LAUNCHERS

PyTorch patched a critical H100 test failure caused by upstream flash-attention directory churn, while Helion merged a major Pallas launcher unification that simplifies the runtime API.

python 20 shipped 1-min read

@pytorch 5 days ago

DYNAMO FIXES ITERATOR CORRUPTION AND FAKE INT ARITHMETIC

Dynamo's deque iteration was silently failing to detect mutations and reversing into the wrong iterator type, while custom hash functions using id() arithmetic crashed the compiler.

python 56 shipped 1-min read

— or follow the whole beat:

Elsewhere on the wire

@cachyos about 9 hours ago

CACHYOS 7.2-RC3-3 LANDS WITH AMD DISPLAY FIXES AND TSC TUNING

CachyOS Linux 7.2-rc3-3 ships critical AMD graphics vblank handling consolidation and direct TSC synchronization for systems without IA32_TSC_ADJUST support.

infra 55 shipped 1-min read

Elixir & Phoenix about 9 hours ago

PHOENIX LIVE VIEW PATCHES WEBSOCKET DOS WHILE ECTO FIXES RELATION TRACKING

Phoenix Live View shipped a critical websocket memory exhaustion fix in its dependency chain, while Ecto quietly resolved a data integrity bug in relationship change detection.

elixir 13 shipped 1-min read

@dotnet about 9 hours ago

CLANGSHARP CLOSES THE C++ AST GAP, MACIOS SHIPS XCODE 26.6 SUPPORT

ClangSharp landed the native bridges and managed API expansion needed to faithfully parse real-world C++ headers, while .NET for macOS is now Xcode 26.6 ready.

dotnet 85 shipped 2-min read

@spring-projects about 9 hours ago

SPRING BOOT CLEANS HOUSE WHILE INTEGRATION FIXES KAFKA BATCH MODE

Spring Boot is removing deprecated code and exposing critical NIO APIs, while Spring Integration ships a fix for Kafka batch processing that was breaking downstream splitters.

java 37 shipped 1-min read