masahi opened a new pull request, #10936: URL: https://github.com/apache/tvm/pull/10936
Per-channel quantization for broadcast ops was supported in https://github.com/apache/tvm/pull/10718, but apparently an invalid `BroadcastAttrs` can be created, where the channel axis can be set to 1 even though the input shape is of rank 1. I've hit this error when working on the quantized BERT model with the latest `main`. The error can be reproduced by running the test (disabled on CI since it is slow) https://github.com/apache/tvm/blob/c2488ac863a8d17ab6b95d67ff19227ed4b2fcbe/tests/python/unittest/test_meta_schedule_integration.py#L220 This PR adds a workaround for cases like that, but I have a feeling that something needs to be fixed on the FQ2I side, where those invalid axes are created. @sfvaroglu @AndrewZhaoLuo @mbrookhart @anwang2009 -- 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]
