comaniac commented on pull request #9782: URL: https://github.com/apache/tvm/pull/9782#issuecomment-998957123
That's an interesting finding. In general, similar to BatchNorm, LayerNorm itself could be converted to FP16 as long as mean and variance stick to FP32, but I don't have this mechanism for now so this is a fair workaround. Meanwhile, do you think ToMixedPrecision should be applied later than the expression mutation passes (e.g., after SimplifyExpr, SimplifyInference, etc, and before FuseOps), so that we could make sure all ops (types) in the IR are basically fixed? -- 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]
