TODO add more context / a repro, for now just capturing the request from chat copied from https://github.com/pytorch-labs/float8_experimental/issues/249