PyTorch: Faster Guards and Kernels, Tighter Supply Chain
Inductor cut compile times and unlocked higher-dimensional tiling while Dynamo fixed correctness gaps and slashed guard overhead. At the same time, release branches removed an unapproved wheel upload and serialization hardened against untrusted imports.
Duration: PT2M39S
Episode overview
This episode is a short developer briefing from PyTorch.
It explains recent repository work in plain language.
- Show: PyTorch
- Published: 2026-11-21T14:00:22Z
- Audio duration: PT2M39S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Hello and welcome, it's Friday, November 21st, 2026.
The lead story is performance with fewer surprises: the compiler stack got faster and more correct, while release engineering and core serialization got safer.
First, Inductor and Dynamo. A large landed change, PR 204324, adds opt-in higher-dimensional tiling for pointwise and reduction kernels. By default nothing changes, but setting max tiles above three now preserves more independent axes, which helps permutes, broadcasts, and fused multiply-sum kernels use hardware…
Second, hardening. Five backports across release branches 2.11 through 2.15, including PRs 205887 and 205891, remove the Flash Attention 3 wheel upload workflow that lacked the approval gate now required on main. For pytree, PRs 205994 and 205993 add a loader that refuses to import modules and reject serialized…
What's next to watch is a large proposed virtual machine stack from contributor bobrenjc93, including PRs 206001 and 206004, to derive graph backward passes in the engine, plus an open fix to upcast 8-bit floats for comparisons, PR 206009, and early work on Apple silicon softmax parity.
That's it for today. Thanks for listening, and happy building.
Nearby episodes from PyTorch
- Stable ABI Hardening and Dynamo Fidelity
- Resize Fixes, Inductor Reliability, and ROCm Gains
- Test Standardization and Compiler Correctness
- Distributed Test Cleanup, Native Fake Tensor Push, and 3.14 Readiness
- C++ Fake Tensor Push and XPU Hardening
- Weekly Recap - Native Fake Tensor Push and Compiler Correctness
- Model Loading Security Fixes and MoE Scaling
- C++ FakeTensor Parity Push