两个self.relative_positions_encoding[:to_seq_length, :to_seq_length, :].to(hidden_states.device)太影响性能了
huangyc0618 opened this issue · comments
huangyc0618 commented
占用了大量CPU资源和时间,建议init初始化后就直接to device
NEZHA: Neural Contextualized Representation for Chinese Language Understanding
huangyc0618 opened this issue · comments
占用了大量CPU资源和时间,建议init初始化后就直接to device