NVIDIA/pytorch-windows-ci

NVIDIA/pytorch-windows-ci

Description: Out of tree repo to run pytorch CI tests on Nvidia GPUs for Windows

Language: Shell

License: NOASSERTION

Stars: 0

Forks: 0

Open issues: 1

Created: 2026-05-22T21:17:24Z

Pushed: 2026-06-30T15:06:38Z

Default branch: main

Fork: no

Archived: no

README:

PyTorch OOT Windows CI

Out-of-tree (OOT) GitHub Actions CI that builds and tests `pytorch/pytorch` on NVIDIA's self-hosted Windows + RTX runners, across multiple Python and CUDA toolkit combinations.

Overview

This repository hosts the GitHub Actions workflows that build and test PyTorch on NVIDIA's self-hosted Windows + RTX runner pool. It implements the downstream half of RFC-0050: Cross-Repository CI Relay for PyTorch Out-of-Tree Backends.

Upstream covers a single configuration (Python 3.12, CUDA 12.8); this repo deliberately expands the matrix to catch regressions across multiple Python and CUDA toolkit combinations before they show up upstream. The build/test logic comes entirely from PyTorch's in-tree .ci/pytorch/*.sh scripts — this repo holds only the workflow wiring. Every job runs on a self-hosted NVIDIA runner; there are no GitHub-hosted (cloud) runs.

> Full architecture, matrix, and runner model: > [docs/ci-details.md](docs/ci-details.md).

Getting Started

The two top-level workflows run automatically on a nightly schedule. No action is required to run them:

• `windows-rtx-build-test.yml` — full source build + test (nightly).
• `windows-rtx-wheel-test.yml` — nightly published-wheel test.

Schedules are documented in [docs/ci-details.md](docs/ci-details.md).

Requirements

• Self-hosted Windows runners from NVIDIA infrastructure, labelled for the

build/test pools (rtx-build, rtx-40x0-test, rtx-50x0-test) and tagged per Python/CUDA cell (py3xx, cu1xx).

• Ephemeral, pre-prepped runner images carrying Python, the CUDA toolkit +

driver, MSVC build tools, and the PyTorch test runtime — the workflows do zero in-job setup. See [docs/ci-details.md](docs/ci-details.md) for the full image contents and label routing.

Usage

The nightly schedules run both top-level workflows automatically. The reusable workflows (_rtx-build.yml, _rtx-test.yml) are called by the two orchestrators and are not run directly.

Detailed reference — workflow table, job naming, install paths, default matrix, test environment variables, and runner diagnostics — is documented in [docs/ci-details.md](docs/ci-details.md).

Performance

Not applicable — this repository provides CI infrastructure rather than a shippable runtime artifact.

Releases & Roadmap

This repo is CI infrastructure and does not publish versioned releases. Changes land via pull request to main.

Contribution Guidelines

Refer to [CONTRIBUTING.md](CONTRIBUTING.md).

Governance & Maintainers

Maintained by the NVIDIA PyTorch Windows CI team. Open an issue or pull request for questions, triage, or proposed changes.

Security

Please report security vulnerabilities responsibly. See [SECURITY.md](SECURITY.md) for the disclosure process. Do not file public issues for security reports.

Support

Maintained on a best-effort basis. For questions, bugs, or feature requests, please open a GitHub issue in this repository.

Community

Discussion happens through GitHub issues and pull requests on this repository.

References

• RFC-0050: Cross-Repository CI Relay for PyTorch Out-of-Tree Backends
• pytorch/pytorch
• [Detailed CI architecture & reference](docs/ci-details.md)

License

This project is licensed under the MIT License — see [LICENSE](LICENSE) for the full text and [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md) for third-party OSS notices.