I. GraphNet: The First Step Toward Next-Generation Compiler

We invite you to become contributors!

Current deep learning models heavily depend on manual kernel optimizations that tightly bind model algorithms and compiler implementations to specific hardware, driving up development costs. AI for Compilers (AI4C) offers a promising alternative by integrating deep learning into tensor compiler backends to accelerate the search for optimal performance. With LLMs advances, we can foresee intelligent compilers that automatically transfer operators-fusion patterns across diverse hardware platforms, achieving true end-to-end optimization. We term this paradigm the “AI Infra Machine Tool”.

To support this vision, we introduce GraphNet, a large-scale collection of computation graphs, intended as a standard dataset for training and validating AI-driven tensor compilers. We aim to collect over one million computation graphs drawn from more than 50 distinct model categories spanning NLP, computer vision, and multimodal domains. To ensure the correctness of our samples, we designed a concise yet powerful set of constraint rules to guarantee the reproducibility of the computation graph extraction process.

We believe that GraphNet will lay the groundwork for a new generation of research in “AI for Systems, Systems for AI.”

By participating in this event, you will learn about the design of GraphNet and AI4C tensor-compiler concept, develop a deeper understanding of the PyTorch and Paddle frameworks, and gain hands-on experience contributing to these open-source deep learning platforms.

II. Task Registration

To prioritize dataset richness and scale, also to avoid duplicate work caused by update delays, developers must pre-register their target tasks (e.g., a specific model on a specific framework):

1. Check for existing registration Google Docs
2. Fill out the registration form
   1. Select the appropriate worksheet by model type: Multimodal, CV, NLP, Speech, RL & Robotics, Others
   2. Provide the following details: Model name, Source platform, Github ID
3. Await GraphNet team review (Please check Task Registration Review column for updates)
   1. Task Registration Review
   2. For expedited review, please @ the team via Discord
4. If approved, you may begin work. If rejected, please revise and resubmit.
5. Once a task is successfully registered, it is locked for one week. If not completed within that week, it becomes available for others to claim.

III. How to Contribute

We provide detailed tutorial paired with tools that help developers to efficiently collect computation graphs. Please see Co-Creation Tutorial or 共创者指引

IV. Acceptance Criteria

Our CI-based validation uses GitHub Actions:

• Pass (green) : Your PR is ready for GraphNet team review. After approval, we will merge it.
• Fail (red) : Update your PR according to the error logs in the “Checks”, submit again.

Core validation aligns with the Validation API to ensure GraphNet dataset construction constraints (check README) are met.

Additional requirements:

• Each PR must contain changes for only one model (graph or script).
• Follow the PR template.
• Do not duplicate previously completed work.