PR Category

User Experience

PR Types

Bug fixes

Description

pcard-85841
in dygraph mode, model will run error: "Incompatible constructor arguments, there are only 5 position args, but the first position args should be PyArray or dtype." when meets 2 conditions:

1. FLAGS_enable_pir_api=true
2. use the following interface to create Tensor (such as recompute)

tmp_tensor = core.eager.Tensor(
                            inner_x.dtype,
                            inner_x.shape,
                            inner_x.name + "cpy",
                            core.VarDesc.VarType.LOD_TENSOR,
                            inner_x.persistable,
                        ) 

The root cause is: In the current situation (default FLAGS_enable_pir_api=false), inner_x.dtype returns paddle.base.libpaddle.VarType object, and after setting FLAGS_deable_pir_mapi=true, it will return paddle.base.libpaddle.DataType object. But core.eager.Tensor requires the data type to be the former, and if the data type is the latter, an error will be reported.

From the original purpose of the PIR design, the data type in the PaddlePaddle framework needs to be unified to the paddle:: DataType (that is the paddle.base.libpaddle.DataType object), so interface core.eager.Tensor should support paddle:: DataType.

Of course, for compatibility, before the official update of PIR(default FLAGS_deable_pir_mapi=True) and during the transitional period, we also supports paddle.base.libpaddle.VarType.