Pytorch with intel gpu

[kukevarius]

I am unable to find how to get pytorch to work with the intel max 1100 gpu.

I dont know what device type i should select in the following code, and when selecting anything else than cuda or cpu it justs says something along the lines of pytorch was not linked/compiled with x

code that i want to run on the intel gpu:

from diffusers import DiffusionPipeline
import torch

pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16, use_safetensors=True, variant="fp16")
pipe.to("cuda")

# if using torch < 2.0
# pipe.enable_xformers_memory_efficient_attention()

prompt = "An astronaut riding a green horse"

images = pipe(prompt=prompt).images[0]

in this case pipe.to("cuda") does not work because no cuda on intel gpu's but xpu hpu opencl all of them dont work + I have no idea which I am even supposed to select to use the gpu.

I am using the free cloud developer service so I do not have root access.
xpu-smi discovery returns all 8 installed intel max 1100 gpu's.

Could you please provide the necessary dependencies and instructions for making this work? Thanks in advance.

[Wan_Intel]

Hi kukevarius,

Thanks for reaching out to us.

To install Intel® Extension for PyTorch on your instance, you may refer to the installation steps at Intel® Extension for PyTorch Installation Guide. Once the installation succeeds, you may visit Intel® Extension for PyTorch Quick Start Guide and Intel® Extension for PyTorch Examples to start using the extension in your code. For more information, please refer to Intel® Extension for PyTorch Documentation.

Meanwhile, could you please provide the following information with us?

• Severity: Low: General guidance/Medium: System impaired/High: System down/Critical: Business-critical system down
• Backup Contact (Name/Number/Timezone):
• System Type: BMaas - Bare Metal / Virtual Machine
• System ID:
• Reservation Start Date and End Date:
• Important Data Stored: Yes / No
• IDC Support Team Can Access The Instance?: Yes / No

Regards,

Wan

[kukevarius]

Hello Wan,

Severity: Low, general guidance

System type: Free Intel Developer Cloud Machine

System ID: ubb083db7da1ae7febbf5cf6a4045f19@idc-training-gpu-compute-04

Important Data Stored: No

IDC Support Team Can Access The Instance: Yes

So i have installed the latest xpu/gpu library with this command,

python -m pip install torch==2.1.0.post2 torchvision==0.16.0.post2 torchaudio==2.1.0.post2 intel-extension-for-pytorch==2.1.30.post0 oneccl_bind_pt==2.1.300+xpu --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/

and i have adapted my code according to the Quick Start Guide,

import torch
from diffusers import DiffusionPipeline
import intel_extension_for_pytorch as ipex

# Load the pipeline
pipe = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    use_safetensors=True,
    variant="fp16"
)

# Move pipeline to Intel Max GPU
pipe.to("xpu")

prompt = "An astronaut riding a green horse"
images = pipe(prompt=prompt).images[0]

I tried this already and I keep getting an error similar to this,

(intel_gpu) ubb083db7da1ae7febbf5cf6a4045f19@idc-training-gpu-compute-04:~$ python intel_demo.py 
/home/ubb083db7da1ae7febbf5cf6a4045f19/intel_gpu/lib/python3.9/site-packages/diffusers/models/transformers/transformer_2d.py:34: FutureWarning: `Transformer2DModelOutput` is deprecated and will be removed in version 1.0.0. Importing `Transformer2DModelOutput` from `diffusers.models.transformer_2d` is deprecated and this will be removed in a future version. Please use `from diffusers.models.modeling_outputs import Transformer2DModelOutput`, instead.
  deprecate("Transformer2DModelOutput", "1.0.0", deprecation_message)
Traceback (most recent call last):
  File "/home/ubb083db7da1ae7febbf5cf6a4045f19/intel_demo.py", line 3, in <module>
    import intel_extension_for_pytorch as ipex
  File "/home/ubb083db7da1ae7febbf5cf6a4045f19/intel_gpu/lib/python3.9/site-packages/intel_extension_for_pytorch/__init__.py", line 95, in <module>
    from .utils._proxy_module import *
  File "/home/ubb083db7da1ae7febbf5cf6a4045f19/intel_gpu/lib/python3.9/site-packages/intel_extension_for_pytorch/utils/_proxy_module.py", line 2, in <module>
    import intel_extension_for_pytorch._C
ImportError: /opt/intel/oneapi/mkl/2025.0/lib/libmkl_sycl_data_fitting.so.4: undefined symbol: _ZN4sycl3_V17handler30verifyUsedKernelBundleInternalENS0_6detail11string_viewE

I do not know how to solve this since i followed the installation guide and i can not find any fixes that work online. 

[Wan_Intel]

Hi kukevarius,

Thanks for sharing the information with us.

Referring to the encountered error: ImportError: /opt/intel/oneapi/mkl/2025.0/lib/libmkl_sycl_data_fitting.so.4 it might be due to an incompatible version of Intel® Extension for PyTorch and Intel® oneAPI Base Toolkit.

For your information, different versions of Intel® Extension for PyTorch require different versions of Intel® oneAPI Base Toolkit. Could you please try the steps below and see if the issue can be resolved?

1. Select Platform (CPU or GPU), Version of Intel® Extension for PyTorch, Operating System, and Package via the following link:
   https://intel.github.io/intel-extension-for-pytorch/#installation?request=platform
2. Follow the installation instructions to re-install Intel® Extension for PyTorch and Intel® oneAPI Base Toolkit

Regards,

Wan

[kukevarius]

Hello wan,

I use the free machine you get from intel so i do not have root access, meaning that i can not change the openAPI Base Toolkit version.

I have tried every single version of Intel extension for pytorch from,

https://intel.github.io/intel-extension-for-pytorch/#installation?request=platform

I selected GPU --> 'i tried all versions' --> Linux/WSL2 --> pip --> generic python.

But every single one of them gives exactly the same error:

(pytorch-8) (/opt/intel/oneapi/intelpython) ubb083db7da1ae7febbf5cf6a4045f19@idc-training-gpu-compute-30:~$ python test.py 
Traceback (most recent call last):
  File "/home/ubb083db7da1ae7febbf5cf6a4045f19/test.py", line 1, in <module>
    import torch
  File "/home/ubb083db7da1ae7febbf5cf6a4045f19/pytorch-8/lib/python3.9/site-packages/torch/__init__.py", line 192, in <module>
    from torch._C import *  # noqa: F403
ImportError: /home/ubb083db7da1ae7febbf5cf6a4045f19/pytorch-8/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so: undefined symbol: iJIT_NotifyEvent

The code that I am using to test if the installations work is as follows:

import torch
from diffusers import DiffusionPipeline
import intel_extension_for_pytorch as ipex

# Load the pipeline
pipe = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    use_safetensors=True,
    variant="fp16"
)

# Move pipeline to Intel Max GPU
pipe.to("xpu")

prompt = "An astronaut riding a green horse"
images = pipe(prompt=prompt).images[0]

This code returns the aforementioned error.

I do not know what I am doing wrong.

[Wan_Intel · Accepted solution]

Hi kukevarius,

Thanks for the information.

We regret to inform you that to use Intel® Extension for PyTorch, it is required to install Intel® oneAPI Base Toolkit.

For your information, Intel® Developer Cloud Virtual Machine Instance does not provide sudo access while Intel® Developer Cloud Bare Metal Instance does provide sudo access.

Sorry for the inconvenience and thank you for your support.

Regards,

Wan

[kukevarius]

Hello Wan,

Thank you for the explanation.

Is there any other way to use the GPU's available on the free cloud machine? (8x Intel max 1100 datacenter gpu, from xpu-smi discovery)

Regards, Kukevarius 

[Wan_Intel]

Hi kukevarius,

For your information, to use Intel® Extension for PyTorch with Intel® GPU, sudo access to install the required packages is needed.

On another note, Intel® Developer Cloud Virtual Machine Instance does not provide sudo access, therefore, it is not possible to use Intel® Extension for PyTorch on Intel® Developer Cloud Virtual Machine Instance with Intel® GPUs.

On the other hand, a workaround would be requesting an Intel® Developer Cloud Bare Metal Instance which provides sudo access.

Regards,

Wan

[kukevarius]

Thank you for the help Wan.

[Wan_Intel]

Hi kukevarius,

If you need additional information from Intel, please submit a new question as this thread will no longer be monitored.

Regards,

Wan