xpu training on laptop with dual NVIDIA T500 and Intel Iris XE cards

[filippo_ciceri]

Hello everyone,

I am trying to run pytorch with gpu support on a laptop with two graphic cards (NVIDIA T500 and Intel Iris XE). I successfully installed intel_pytorch_extension following this guide.

The xpu device is recognized (I get "True" from torch.xpu.is_available()) and I can transfer tensors and execute simple operations such as sums, multiplications, etc... If I try to perform inference with a LLM (GPT2), or simply cloning a tensor, the process takes forever and I eventually get this error:

RuntimeError: Native API failed. Native API returns: -999 (Unknown PI error) -999 (Unknown PI error)

On task manager, I see that memory is allocated on the slower Intel Iris Xe, but no processing is actually going on. No memory is allocated and no processing is performed on the NVIDIA T500.

Is there a way to select the NVIDIA T500 for this task?

Do you know how to avoid getting this -999 error?

thanks

[asin · Accepted solution]

Hi Filippo,

Intel Extension for PyTorch for XPU is only supported on the following hardware:

• Intel® Data Center GPU Flex Series 170
• Intel® Data Center GPU Max Series
• Intel® Arc™ A-Series GPUs

It is not supported on your integrated GPU on your laptop. That is why you are getting the Native API failed error. To select your T500, you would set your torch device to "cuda". Let me know how else I can help.

Best,

Alex

[filippo_ciceri]

Thank you for the quick reply.

This is indeed solved the issue! 

Filippo