tlopex commented on PR #20135: URL: https://github.com/apache/tvm/pull/20135#issuecomment-5466506433
For a float32 base, `rhs` is converted to float64 above but then rounded back to float32 by `relax.const(rhs, power_dtype)`. e.g. `2**24 + 1` becomes `2**24`, which can change the magnitude for nontrivial inputs. Could we evaluate float32 power in float64 and cast back, and add a regression test that does not only use `-1`? -- 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] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
