50% Drop in Windows Issues Moves PyTorch Toward Cross-Platform Stability
PyTorch
Disney used it to achieve a 10× speed-up in animated video processing; Duolingo increased user retention by 12%; and Lyft cut model training from days to one hour.
PyTorch, an open-source deep learning framework originally developed by Meta, powers breakthroughs like these across natural language processing, computer vision, speech, healthcare imaging, autonomous systems, scientific research, and more.
Unfortunately, Windows developers couldn’t utilize the framework’s full power because historically, PyTorch on Windows suffered from a significantly higher number of issues than on other platforms.
Industry: Open Source, AI/ML Big Data
Technologies: C & C++, C# & .NET, Windows
Solutions: Open Source, Porting, Software Maintainer
The Challenge
Janea Systems joined forces with PyTorch to close the cross-platform stability gap so that Windows developers have the same experience and access to the same capabilities as those working with PyTorch on Linux and other platforms.
At project start there were 187 open Windows-specific issues versus just over half that number on other platforms. Our objective was to drive Windows issues down to parity range, so that Windows developers can fully harness PyTorch’s true power.
Client impact
We ran a focused fix effort to push PyTorch towards parity by strengthening tests, adding key performance tools, modernizing builds, and clearing compatibility blocks:
- Systematic Test Restoration & Hardening: Audited, fixed, and re-enabled tests to raise real coverage on Windows.
- ARM64 Enablement via Performance Libraries: Integrated and configured required ARM performance libraries.
- Dynamic Linking & Expanded Runtime Support: Switched Intel MKL linking from static to dynamic and added CUPTI integration for Kineto profiling.
- Comprehensive Refactoring for C++20 Modules: Reworked the codebase for smooth integration of MSVC C++20 modules.
We eliminated noise, closed critical gaps, and removed blockers, so PyTorch on Windows now operates on par with the other platforms. The result: significantly fewer open issues, full profiling and optimization tooling, cleaner modern C++ builds, and broader heterogeneous hardware support.
If your team is facing platform-specific challenges or needs to stabilize complex, multi-platform systems, we can help. Improved developer experience, advanced profiling, or better performance on Windows and beyond: let’s talk about how Janea Systems can support your engineering goals.
The Team
Other Client Successes
Let's talk about your project
Let's talk about your project
Ready to discuss your software engineering needs with our team of experts?