PyTorch: Weekly Recap - Reverts and Infrastructure Updates
PyTorch had an unusual week with zero merged pull requests and 30 commits, including multiple automatic reverts of recently added features and significant infrastructure updates moving CI jobs to CUDA 13.0.
Duration: PT2M19S
Episode overview
This episode is a short developer briefing from PyTorch.
It explains recent repository work in plain language.
- Show: PyTorch
- Published: 2026-04-27T00:00:00Z
- Audio duration: PT2M19S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Good morning. This is your PyTorch weekly recap for April 20th through 27th, 2026.
Zero pull requests merged and 30 additional commits this week - an unusual pattern dominated by automatic reverts and infrastructure changes.
Starting with fixes, Avi resolved a critical dtype promotion bug in CUDA max and min kernels. The issue caused torch.maximum and torch.minimum to return silently incorrect results when using type promotion with narrower output tensors. The fix corrected four kernel functions to use common_dtype instead of output…
The week saw significant revert activity from PyTorch's automatic revert system. Three major features were rolled back: a naive linear cross entropy implementation, the corresponding LinearCrossEntropyLoss module, and inductor materialization heuristic improvements for computation chains. These reverts removed…
On infrastructure, Ting Lu completed a substantial migration moving all CUDA 12.8 CI jobs to CUDA 13.0, affecting over 20 workflow files. This aligns with CUDA 13.0 becoming the stable release for PyTorch 2.11.
Bin Bao enhanced AOTI's constant buffer handling to support mixed-device constants. The previous implementation rejected constants on…
Nearby episodes from PyTorch
- Weekly Recap - Code Quality & Infrastructure Improvements
- Weekly Recap - Dynamo Enhancements & Critical Fixes
- Optimization Engine Revamp and Hash System Breakthrough
- Operator Improvements and Build System Updates
- Inductor Improvements and Bug Fixes
- Optimization Improvements and Cache Unification
- Weekly Recap - Performance Optimizations and Code Cleanup
- Weekly Recap - Infrastructure Stability & Mixed Precision