junrushao1994 opened a new pull request, #11725: URL: https://github.com/apache/tvm/pull/11725
This PR fixes a bug in the pass Narrow-Datatype in TIR, where dtype of certain IterVar and loop variables are adjusted to narrower ones. The bug occurs when the dtype of thread axis is int32, while its extent is int64, where the original behavior will not narrow the extent to int32, which causes an assertion thrown in IterVar's constructor. An alternative approach is to re-dtype IterVar to int64, however, the subsequent passes do not actually respect int64 thread axes, which leads to even more issues in lowering. This bug prevents AutoTIR in tuning Huggingface DistilBERT. -- 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]
