torch::cuda::is_available() is always False in third party plugin in Adobe Premiere pro

[Irfan5EBE]

Hi, I am developing a plugin for Adobe premiere Pro. I used static lib torch in my application. However when I am trying to integrate my application to Adobe premiere pro It is reporting low-level exception error.
When I debugged for the root cause, I observed that torch::cuda::is_available() returns FALSE always. Build Environment for the application is as below.
- VS 2019
- NVIDIA GeForce RTX 3060 and CUDA Tool kit version 11.1
- Pytorch 1.8.1

I can see device_count as 1 with NVIDIA API as below in the plugin solution
result = nvmlDeviceGetCount(&device_count);

But torch::cuda::device_count() returns 0.

Could anyone help why torch is not linked to cuda in the application.

[Bruce Bullis]

I discussed this with the PPro team's GPU specialists, who recommend you not use torch.

There have been significant stability issues, it's only performant on specific Nvidia GPU/Intel CPU combos.

Also, it's not something Adobe actively supports, or will support in the future.

[Irfan5EBE]

Thank you for the reply.

Could you please suggest any other alternatives for torch (such as tensorflow?) which is supported by Adobe premiere pro and can be used for plugin development for NVIDIA GeForce RTX 3060.

[Bruce Bullis]

> Could you please suggest any other alternatives for torch (such as tensorflow?)

Specifically, what are you trying to do, for which torch seems like the correct option?

[Irfan5EBE]

I am trying to integrate one DNN framework/model to Adobe Premiere pro. 

[Bruce Bullis]

Cool...what workflow(s) are you trying to support?

[Irfan5EBE · Accepted solution]

Sorry for the late reply. As it is confidential, I may not be able to disclose it in this channel. Thank you