OmarAzizi commented on issue #19971: URL: https://github.com/apache/tvm/issues/19971#issuecomment-4927917187
Hey @viiccwen, I was digging into this before seeing your comment, and it looks like the root cause is in `Expand`'s `ShapeExpr` broadcasting fast-path at [onnx_frontend.py:2864](https://github.com/apache/tvm/blob/main/python/tvm/relax/frontend/onnx/onnx_frontend.py#L2864). When the target shape needs rank-padding, the no-op shortcut returns the unpadded input instead of broadcasting it up, so two sibling branches end up at different ranks before hitting `Concat`. I've already got a fix and tests passing locally (verified against the model from this issue and the full ONNX frontend suite). Happy to open the PR now, or if you'd rather take it from here. I can also post the diff/analysis for you to use. Let me know which you'd prefer. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
