This repository is designed to allow out-of-tree compute platforms (accelerators) seeking to be listed on the PyTorch website's Additional Compute Platforms page to submit their applications. If you have any questions, please open an issue in this repository.
For the detailed admission criteria, please review the PyTorch Foundation Additional Compute Platforms Admission Process to ensure that you meet the minimum expectations before applying.
Applying is as easy as 1-2-3!
- Open a new Additional Compute Platform Application using the linked issue form.
- Complete all sections, providing evidence (links, docs, CI dashboards, policies) for each requirement.
- Submit the issue.
Once you have submitted your issue, your application is placed in the backlog for triaging by the Accelerator Integration Working Group.
The Co-chair of the Accelerator Integration Working Group (or a delegate designated by the Co-chair) will perform a pre-review of your submission against the Admission Requirements. If gaps are identified, you'll be asked to address them via comments on the issue — be sure to check back periodically. Once the pre-review passes, the Co-chair initiates a vote (see Voting below).
Removing, updating, and adding labels is the responsibility of the Accelerator Integration Working Group. Please refrain from adding additional labels.
The application is new and placed in the backlog awaiting pre-review.
The Working Group has identified gaps against the Admission Requirements and is waiting on the submitter to provide additional information or evidence as a comment on the issue.
The pre-review passed and the subsequent vote succeeded. The platform will be added to the Additional Compute Platforms page.
The vote did not pass, or the submitter did not resolve outstanding gaps. The reasons will be provided in the comments of the issue.
Once an application passes pre-review, the Co-chair initiates a vote. Voting members consist of the Working Group Co-chairs and the TAC members who mentor the Working Group.
A quorum requires more than half of the Working Group's eligible voting members to participate; the vote then passes by a simple majority of those who participate.
Once voting is complete, the issue is labeled and updated with the outcome. Approved platforms receive the benefits described below; declined platforms receive feedback explaining the reasons in the issue comments.
A platform that has been declined, or that is later retired, may reapply via the standard submission process once the issues that led to the decline or retirement have been resolved.
Approved compute platforms receive:
- Listing on the PyTorch Additional Compute Platforms page as a recognized platform backend.
- A blog post announcing the new platform listing from the PyTorch marketing team.
- Collaboration opportunities with the Accelerator Integration Working Group and the broader PyTorch ecosystem.