[feature request] use torch namespace instead of at about extension-cpp HOT 4 CLOSED

checking with @gchanan and @yf225 , is this fine to do?

from extension-cpp.

Yes we should use torch:: in the cpp and cuda modules as we are deprecating the at:: namespace in the public API. @ClementPinard We would love to have PR contribution on it. Thanks!

from extension-cpp.

which cpu and cuda modules are you talking about?

from extension-cpp.

https://github.com/pytorch/extension-cpp/tree/master/cpp
https://github.com/pytorch/extension-cpp/tree/master/cuda

Also, I would like to make this example more up to date in general to help people develop cuda modules easily.
More specificly, I'd like to solve the fmax issue that seems to bother many people since the beginning, and seems to be solvable by a scalar_t cast. (see #27 , this does not seem to be an environment error).

Besides, as suggested #12 , I'd like to add usage of .packed_accessor in the cuda kernel.

Finally, I'd like to update the tutorial code here to reflect those changes, along maybe a mention to packed_accessor in the tensor basics doc here

How does that sound to you ? Unfortunately, i'm stuck at the moment because of #27 but hopefully I'll find a solution or a working setup where the code compiles.

from extension-cpp.