Unable to infer ltx-video on iGPU

[rainhenry]

The path of the file in question:
notebooks/ltx-video/ltx-video.ipynb

I modified some of the code as follows:
##compiled_transformer = core.compile_model(TRANSFORMER_OV_PATH, device.value) ## Original code
compiled_transformer = core.compile_model(TRANSFORMER_OV_PATH, "GPU") ## I modified
compiled_vae = core.compile_model(VAE_DECODER_PATH, device.value)
compiled_text_encoder = core.compile_model(TEXT_ENCODER_PATH, device.value)

My idea is to let the transformer process perform inference on the iGPU, because this part is the most time-consuming part of the entire inference process. The value of device.value in the original code is "AUTO". By observing the load of CPU and iGPU, it is not difficult to find that when it is set to "AUTO", it works on the CPU instead of iGPU. So I manually changed this value to "GPU". After I modified it, the whole inference process seemed to be smooth, and no errors occurred. However, the final generated video file was a pair of chaotic and blurred noise, not a normal video. If I change this value back to CPU, then the generated video file is correct.

So I don't know what the problem is? Is it that I modified it in the wrong way? Or is there some bug in the code?

According to my previous experience of deploying some large models for generating videos to iGPU, when exporting from the original model to IR format, the exported IR model is strongly related to the resolution and frame rate information that needs to be output. I didn't see it in this demo, and I'm not sure if it is related to this reason.

My hardware device is Core Ultra7 155H

[YuChern-Intel]

I noticed that the tutorial only shows "CPU" and "AUTO" as the available devices. In this case, selecting "AUTO" will be the same as selecting "CPU".

There is an added quantization steps in the latest branch compared to 2025.0 which helps in making model inference faster.

[rainhenry]

I noticed that the tutorial only shows "CPU" and "AUTO" as the available devices. In this case, selecting "AUTO" will be the same as selecting "CPU".

There is an added quantization steps in the latest branch compared to 2025.0 which helps in making model inference faster.

You are right. The results of selecting AUTO and CPU are the same. But I want to run the inference process on iGPU. How should I modify it? Because the inference speed of CPU is too slow.

[openvino-dev-samples]

hi @aleksandr-mokrov, as the author of this notebook, do you know this issue on GPU ? thanks

[aleksandr-mokrov]

I've reproduced the problem. Looks like we have an issue on GPU plugin side. I will let you know when the problem is fixed.