PyTorch: Model Loading Security Fixes and MoE Scaling

Four fixes harden checkpoint loading against code execution and path traversal, while Token Switch and NCCL changes cut memory use and restore overlap for mixture-of-experts training. Compiler patches also fix silent correctness bugs in functionalization and Inductor.

Duration: PT2M33S

Episode overview

This episode is a short developer briefing from PyTorch.

It explains recent repository work in plain language.

  • Show: PyTorch
  • Published: 2026-11-15T14:00:31Z
  • Audio duration: PT2M33S

Transcript excerpt

This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.

Hello, it's Sunday, November 15th, 2026 — welcome to your PyTorch daily briefing.

The clearest pattern today is trust: four security fixes close code execution and file escape paths when loading checkpoints, alongside a push to make large mixture-of-experts training faster and leaner.

First, loading untrusted files. Two changes tighten the weights-only unpickler, tracked in PR 205011 and 205013. One blocks direct construction of raw storage objects that could expose uninitialized process memory as a tensor, the other broadens blocked modules and adds function-level blocking for risky calls.…

Second, distributed experts at scale. A Token Switch stack, including PR 204974 and 204973, exposes quantized dispatch and combine with compiler-visible operators and backward support for B-F-16 and M-X-F-P-8. PR 204975 keeps asynchronous exchange tensors in the work object instead of pinning them to the stream,…

Finally, compiler correctness. PR 205026 fixes functionalization zeroing untouched elements when mutating through gapped unfold views, while PR 205018 and 205019 fix Inductor crashes and extra copies around auto-functionalized mutable ops. PR 205025 aligns fake tensor dtype rules for…

What…

Nearby episodes from PyTorch

  1. Test Standardization and Compiler Correctness
  2. Distributed Test Cleanup, Native Fake Tensor Push, and 3.14 Readiness
  3. C++ Fake Tensor Push and XPU Hardening
  4. Weekly Recap - Native Fake Tensor Push and Compiler Correctness
  5. C++ FakeTensor Parity Push
  6. Stricter Export Rules and Edge-Case Hardening
  7. Compiler Correctness Fixes and Accelerator Generalization
  8. Rust Compiler Prototype and Private Device Runtime