is it possible to make a amd version using HSA COMPUTE with OpenCL or a directx12 version of this

[Cxsmo-ai]

is it possible to make a amd version using HSA COMPUTE with OpenCL or a directx12 version of this

[GarrettConway]

Are you not able to use CPU inference?

[Cxsmo-ai]

no want the speed of gpu not cpu

[Cxsmo-ai]

and if it had HSA COMPUTE with OpenCL it would run good on amd and if that's not doable then at least a directx12 version.

[GarrettConway]

You are welcome to attempt to use the PyTorch ROCm build (it may be as simple as just installing this as your torch, but I have no experience with ROCm build), however it only runs on Linux. Unless another contributor comes in to do this though, I seriously doubt any existing contributor will attempt to explicitly add support as I believe we likely do not have AMD equipment to test this and are focusing on other areas.

[GarrettConway]

Pytorch ROCm install can be found here: https://pytorch.org/get-started/locally/
You can try just forcing that as your torch version.
Theoretically, the python code is mostly agnostic about ROCm, so it might work by doing just that or some minimal other changes: https://github.com/pytorch/pytorch/blob/master/docs/source/notes/hip.rst#hip-interfaces-reuse-the-cuda-interfaces

[Cxsmo-ai]

what about this https://github.com/microsoft/antares

[34j]

Unfortunately, there is no need for this repository and it will not be implemented permanently, as the training will only take a few hours and it needs to be compatible with the upstream repository; no PR will be accepted.

[34j]

Migrate to #120