[Prim][PIR] Support the dynamic shape for layer_norm_grad #68274
Conversation
zeroRains
requested review from
Aurelius84,
cxxly,
xiaoguoguo626807,
changeyoung98 and
cyber-pioneer
as code owners
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
|
|
|
| if (bias_ptr->dtype() == phi::DataType::FLOAT16 || | ||
| bias_ptr->dtype() == phi::DataType::BFLOAT16) { | ||
| bias_grad_tmp = cast<T>(bias_grad_tmp, bias_ptr->dtype()); | ||
| } |
There was a problem hiding this comment.
这里的判断逻辑有一个函数ConvertoMT可以使用
There was a problem hiding this comment.
不行吧，我看ConvertoMT里面用是将半精度转成float32，但是这个部分是从float32转回半精度
| } | ||
| set_output<T>(bias_grad_tmp, bias_grad); | ||
| } else { | ||
| bias_grad = nullptr; |
There was a problem hiding this comment.
动静shape 相同的代码只写一份吧
| @@ -49,6 +49,7 @@ | |||
| "pd_op.gather_nd", | |||
| "pd_op.gelu", | |||
| "pd_op.hardswish", | |||
| "pd_op.layer_norm", | |||
There was a problem hiding this comment.
反向动态shape 基本支持完备了，可以看看怎么删除这个list吧
|
Sorry to inform you that 6b46c2b's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually. |
PR Category
Operator Mechanism
PR Types
New features
Description
为layer_norm_grad添加动态shape支持，Mac和cpu精度不足