jwfromm opened a new pull request, #13289: URL: https://github.com/apache/tvm/pull/13289
Due to the maximum value of FP16 being quite low (65536), it's very common to be unable to express large reductions using it. This occurs most commonly in simple operators like `mean` and `var`, where the sum of elements is used as a divisor. Given that the actual complexity of those operators is quite low, it makes sense to just perform them in fp32 instead of fp16. This PR adds that casting and a few tests that previously would fail in FP16. -- 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]
