ipex-llm icon indicating copy to clipboard operation
ipex-llm copied to clipboard

Docker image (intelanalytics/ipex-llm-xpu): Documentation stated I would need to disable iGPU to use A770. When will you fix this issue since disabling iGPU is problematic?

Open sungkim11 opened this issue 1 year ago • 6 comments

Docker image (intelanalytics/ipex-llm-xpu): Documentation stated I would need to disable iGPU to use A770. When will you fix this issue since disabling iGPU is problematic?

sungkim11 avatar May 06 '24 19:05 sungkim11

What are you using it for? You might have some luck with setting this environment variable (after the setvars.sh and source ipex-llm-init lines):

export ONEAPI_DEVICE_SELECTOR=level_zero:0

Set it to level_zero:0 or level_zero:1, see if one of them works. I used this strategy when I was experimenting with dual A770s, one on x4 PCIE and one on x16.

More documentation here:

https://intel.github.io/llvm-docs/EnvironmentVariables.html

digitalscream avatar May 06 '24 20:05 digitalscream

Hi @sungkim11 are you following this document: https://ipex-llm.readthedocs.io/en/latest/doc/LLM/Quickstart/docker_windows_gpu.html# to install IPEX-LLM in Docker, we will try to reproduce from our side and try the method suggested by @digitalscream.

liu-shaojun avatar May 07 '24 01:05 liu-shaojun

Yes, I have followed the document. I am referencing the last section where it suggest I disable iGPU. This is problematic for me.

sungkim11 avatar May 07 '24 04:05 sungkim11

Hi @sungkim11

I tried running a Docker container (intelanalytics/ipex-llm-xpu) on a machine equipped with an A770 processor and integrated graphics (iGPU) following the steps in the documentation. After disabling the iGPU in the device manager, the chat service could start normally.

Although we also attempted method recommended by @digitalscream and method mentioned in the issue: https://github.com/intel/intel-extension-for-pytorch/issues/508,

export ZE_AFFINITY_MASK=0
or
export ONEAPI_DEVICE_SELECTOR=level_zero:1

it seems they all resulted in issues related to RuntimeError: could not create a primitive. https://github.com/intel/intel-extension-for-pytorch/issues/508#issuecomment-2097713733

Given this situation, I'd suggest sticking with the manual method of disabling the iGPU as outlined in the documentation. We'll update you in this issue if we discover any alternative solutions in the future. Apologies for any inconvenience caused.

liu-shaojun avatar May 07 '24 08:05 liu-shaojun

This is problematic since my monitors are using iGPU, not A770. I would lose display when I disable iGPU.

sungkim11 avatar May 07 '24 12:05 sungkim11

We are working on a fix, and we'll update this issue when it's fixed. Alternatively, you can try using A770 for graphical interface support.

liu-shaojun avatar May 08 '24 01:05 liu-shaojun