microsoft / microsoft/onnxruntime
[Feature Request] Enable interop between model editor APIs and new ORT compile APIs
- Dominant language
- C++
- Stars
- 21.9k
- Forks
- 4.2k
- Avg merge
- 4d 11h
- Merged PRs (30d)
- 184
Description
### Describe the feature request
There were some new APIs introduced earlier in the year to streamline the process of compiling ORT models, i.e., https://onnxruntime.ai/docs/api/c/struct_ort_compile_api.html.
These APIs can accept the model as a file path or as a buffer. This feature request proposes extending the API to support models generated via the model editor APIs. https://onnxruntime.ai/docs/api/c/struct_ort_model_editor_api.html#details
### Describe scenario use case
This would be useful for WebNN scenarios where the browser crafts the model graph using the editor APIs. There is currently no way to pass the composed graph API to the compile API. Ideally this could be done in a way that performs the compilation in-memory (the browser's GPU sandbox process is heavily restricted and cannot write data to the file system).
Contributor guide
Research direction
Start with the Ort Compile API and Ort Model Editor API documentation linked in the issue, then trace their corresponding implementation and tests in ONNX Runtime. Define how an editor-created graph can be passed to compilation without filesystem access, and consider the WebNN browser sandbox scenario. Done means the composed graph is accepted by the compile API and compilation can occur in memory.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp
- Domain
- backend-api-design, machine-learning
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 25/100