llnl / acrotensor Goto Github PK

Hi all,

I'm considering trying to integrate acrotensor with PyTorch to use it in a model. It looks like it should be fairly trivial to create an acro::Tensor that references the GPU memory of some_pytorch_tensor.contiguous(); the only issue is Tensor.cpp:L192:

if (ddata != nullptr) {
      ACROBATIC_ASSERT(hdata != nullptr, "Acrotensor does not currently support GPU only tensors.");

It seems to me reading the rest of Tensor.cpp that this isn't actually true — it seems like as long as you never try to move the Tensor back to host it should be fine — but I was hoping to confirm this.

If anyone has previously tried to integrate this with PyTorch or has any thoughts on that, I'd also be grateful to hear them.

Thanks for your work on this library!