comaniac commented on pull request #8108: URL: https://github.com/apache/tvm/pull/8108#issuecomment-846302151
FYI: It's possible that this error is caused by the CUBLAS math flag: https://github.com/apache/tvm/blob/813136401a11a49d6c15e6013c34dd822a5c4ff6/src/runtime/contrib/cublas/cublas.cc#L38 According to the CUBLAS document, this flag is being deprecated. We also made some tests previously around this flag, and found that it is effective even for float32, meaning that CUBLAS kernel internally casts float32 to float16, does the computation, and casts the results back. As a result, this flag may introduce accuracy issue. -- 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. For queries about this service, please contact Infrastructure at: [email protected]
