Remove unnecessary tensor copy in load op

[kexinzhao]

DeserializeFromStream function calls TensorFromStream, where it has the following code:

Paddle/paddle/fluid/framework/tensor_util.cc

Lines 319 to 336 in ccc594e

	if (platform::is_gpu_place(dev_ctx.GetPlace())) {
	#ifdef PADDLE_WITH_CUDA
	Tensor cpu_tensor;
	cpu_tensor.Resize(framework::make_ddim(dims));
	framework::VisitDataType(
	desc.data_type(),
	DeserializedDataFunctor(&buf, &cpu_tensor, ctx.GetPlace()));
	is.read(static_cast<char*>(buf), cpu_tensor.memory_size());
	auto dst_place = dev_ctx.GetPlace();
	framework::TensorCopy(cpu_tensor, dst_place, dev_ctx, tensor);
	#else
	PADDLE_THROW("Unexpected branch");
	#endif
	} else {
	framework::VisitDataType(
	desc.data_type(),
	DeserializedDataFunctor(&buf, tensor, ctx.GetPlace()));
	is.read(static_cast<char*>(buf), tensor->memory_size());

This means that TensorFromStream will first load tensor from disk to CPU place, if the load op is run on GPU place, it will copy the tensor to GPU via framework::TensorCopy(cpu_tensor, dst_place, dev_ctx, tensor);

So in the load op, we don't need to do the copy from CPU to GPU again.

[reyoung]

Excellent