LLM cannot run on NPU devices.

[ZhangGuoLong666]

Why does my Python program suddenly end when calling the LLMPipeline (model_cath, 'NPU') method when I run the LLM model on OpenVino.genai and choose to run it on NPU. Console printing: Process finished with exit code -1073741819 (0xC0000005)

[Wovchena]

It could be out of memory error.

1. What's the optimum-cli command to obtain the model?
2. Does it work on CPU?

[ZhangGuoLong666]

It could be out of memory error.

1. What's the optimum-cli command to obtain the model?
2. Does it work on CPU?

I have 64GB of memory and am running a 1.5B model. Based on my AIPC, there is still a lot of memory usage left. If I set the method parameters to CPU or GPU, it can run normally. I have tested multiple models, including Qwen3-1.5B and Deepseek-R1-Qwen3-7B, each of which can use GPU or CPU. And I also tried running it on another AIPC with the same result. I tried switching between openvino. genai versions 2025.2.0.0 and 2025.3.0.0, and the same situation occurred. My option cli is using the instructions from Intel's official tutorial: optimum-cli export openvino --model Qwen/Qwen3-8B --task text-generation-with-past --weight-format int4 --group-size 128 --ratio 0.8 Qwen3-8B-int4-ov

[Wovchena]

I assigned to NPU. At the meantime you can try upgrading NPU driver: https://docs.openvino.ai/2025/openvino-workflow/running-inference/inference-devices-and-modes/npu-device.html

[ZhangGuoLong666]

I assigned to NPU. At the meantime you can try upgrading NPU driver: https://docs.openvino.ai/2025/openvino-workflow/running-inference/inference-devices-and-modes/npu-device.html

Okay, I will come to the company tomorrow to try the method you gave me and see if it can solve the problem. Thank you for your reply. I will tell you the result tomorrow.