Recently I have found that as the amount of communication increases in torch.Tensor.cpu(), the bus bandwidth (is calculated by amount of communication / latency) will increase accordingly, but when the amount of communication is 16MB, there will be a sudden drop in bus bandwidth (the latency increases largely), so I want to see the implementation. But where can I find the source code of torch.Tensor.cpu()
The document only describe the function of torch.Tensor.cpu, but this document doesn’t give the source code of this method.
True, couldn’t find it either.
Actually I think I find some source code related to this function:pytorch/aten/src/ATen/core/Tensor.cpp at main · pytorch/pytorch · GitHub. This function will be called when calling cpu(), but this function will call another function that may be generated by code generation. So we cannot view the source code.
1 Like