【Hackathon 6th No.55】move fake_quantize_dequantize_moving_average_abs_max/fake_quantize_dequantize_abs_max op to phi #63737
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
感觉应该是梯度计算出问题了，我试过将fluid中的方法改写成phi的格式，并将functor迁移到phi下，是可以通过CI中报错的单测的，但是将整个op迁移到phi下就产生了错误，需要老师帮忙看一下怎么解决。 |
| self.check_grad( | ||
| ['X'], 'Out', user_defined_grads=gradient, check_dygraph=False | ||
| ) |
There was a problem hiding this comment.
感觉和这里有点关系，因为还在fluid下的op测试是不需要设置check_dygraph=False的选项的，在迁移到phi下会出现下图的问题，我是不是漏配置了什么东西，也需要老师帮忙看一下。
There was a problem hiding this comment.
ci上看到这个单测通过了，你重新提交一个commit过下所有CI再观察下
| _C_ops.assign_out_(out2, self._scale) | ||
| _C_ops.assign_out_(out3, state) | ||
| _C_ops.assign_out_(out4, accum) | ||
| return quant_out |
There was a problem hiding this comment.
按照提供的解决方案给对应原本ouput的变量进行赋值，但是由于原本的self._scale是一个可选参数，可能导致没有初始化，从而产生了这个问题，请问应该如何修改呢？
There was a problem hiding this comment.
@YuanRisheng 麻烦老师帮忙看下_(:з」∠)_
There was a problem hiding this comment.
试试用_is_initialized方法先判断一下out，再选择看是否assign
There was a problem hiding this comment.
emmm，如果能进入这个分支应该已经init了，但是还是报错，(:з」∠)
There was a problem hiding this comment.
试试用_is_initialized方法先判断一下out，再选择看是否assign
判断out2，不是判断_scale
|
CI通过了~ |
…_max/fake_quantize_dequantize_moving_average_abs_max op to phi - part (PaddlePaddle#63737) * move functor to phi * fix namespace * modify to phi method and fix the bug in dequantize_log op when set the pir * move fake_quantize_dequantize_moving_average_abs_max to phi * fix bug * move fake_quantize_dequantize_abs_max to phi * roback * modify the dynamic graph function call * fix dynamic modify * still have bug * add the assign * roback * merge * merge * fix bug * fix merge bug * fix bug * fix bug * fix bug
PR Category
Others
PR Types
Others
Description
需先Merge前置PR：#63732
move fake_quantize_dequantize_moving_average_abs_max to phi