Repository navigation
Proposal: a browser model diagnostics workbench with exportable reproductions #33737
Replies: 1 comment
|
This is a fantastic proposal. One of the biggest friction points when triaging browser-based ML issues in Here are concrete architecture thoughts on where this fits within the ONNX Runtime codebase and the key diagnostics such a workbench should capture: 1. Where This Fits Best:
|
Uh oh!
There was an error while loading. Please reload this page.
I would like to build a small workbench around ONNX Runtime Web that turns a model-loading or execution failure into a useful reproduction package. I would reuse the existing WebGPU/WASM paths and diagnostics.
My first version would load one model, record browser/backend capabilities and runtime versions, attempt supported execution paths, and explain which stage failed. An export would contain sanitized configuration, logs, and the minimal harness, with model inclusion explicitly chosen by the user. I would separate observed execution results from static compatibility guesses.
Would this fit the Web examples/tooling area, or would you prefer a companion repository? I can implement the first harness and report format after agreeing which diagnostics are stable enough to expose.
I have taken a similar proposal through upstream discussion and implementation in Qwen Code: the Android direction was accepted, followed by my merged shell/PWA contribution.
All reactions