[Eager] fix set_value logic when input's dtype is different #48519
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
paddle/fluid/pybind/eager_method.cc
Outdated
| @@ -1243,6 +1242,10 @@ static PyObject* tensor_method__setitem_eager_tensor(TensorObject* self, | |||
| value_tensor = egr::EagerAmpAutoCast( | |||
| value_tensor.name(), value_tensor, amp_dtype, "set_value"); | |||
| } | |||
| if (value_tensor.initialized() && | |||
There was a problem hiding this comment.
why not merge this into branch above.
There was a problem hiding this comment.
merge this into branch above now.
paddle/fluid/pybind/eager_method.cc
Outdated
| @@ -1243,6 +1242,10 @@ static PyObject* tensor_method__setitem_eager_tensor(TensorObject* self, | |||
| value_tensor = egr::EagerAmpAutoCast( | |||
| value_tensor.name(), value_tensor, amp_dtype, "set_value"); | |||
| } | |||
| if (value_tensor.initialized() && | |||
| self->tensor.dtype() != value_tensor.dtype()) { | |||
| value_tensor = cast_ad_func(value_tensor, self->tensor.dtype()); | |||
There was a problem hiding this comment.
This is not a safe operants， check dtype first maybe
There was a problem hiding this comment.
Checkout the EagerAmpAutoCast && GetAmpDestDtype system
There was a problem hiding this comment.
后续的set_value__dygraph_function的两个输入必须相同dtype, 而先前的 EagerAmpAutoCast 并不能保证两个输入相同 dtype, 比如tensor A float16, tensor B float64, 经过 EagerAmpAutoCast 处理，tensor A 为 float32, 而 tensor B 依然 float64，调用 set_value__dygraph_function 会报错的。所以需要调用 cast_ad_func 进行对齐。
self->tensor = set_value__dygraph_function(
self->tensor, value_tensor, {}, {}, {}, attrs);
…ddle#48519) * [Eager] fix set_value logic when input's dtype is different * value_tensor * fix set_value logic when input's dtype is different
PR types
Bug fixes
PR changes
Others
Describe
fix set_value logic when input's dtype is different