[Bug]: docker image openvino/model_server:latest-gpu does not serve the model correctly

[fedecompa]

OpenVINO Version

2024.3

Operating System

Windows System

Device used for inference

intel UHD Graphics GPU

Framework

None

Model used

meta-llama/Llama-3.2-3B-Instruct

Issue description

I deployed the llama 3.2 -3B model using the image: openvino/model_server:latest-gpu following the documentation here:

https://docs.openvino.ai/2024/openvino-workflow/model-server/ovms_demos_continuous_batching.html

and the folder structure for the openvino IR model:

https://github.com/openvinotoolkit/model_server/blob/main/docs/models_repository.md

The command in my docker-compose is:
command: --model_path /workspace/Llama-3.2-3B-Instruct --model_name meta-llama/Llama-3.2-3B-Instruct --port 9001 --rest_port 8001 --target_device GPU

From the logs in the container I see that the server loads the model and starts correctly. Indeed if I call the API http://localhost:8001/v1/config I obtain:

{
"meta-llama/Llama-3.2-3B-Instruct" :
{
"model_version_status": [
{
"version": "1",
"state": "AVAILABLE",
"status": {
"error_code": "OK",
"error_message": "OK"
}
}
]
}
}

However when I call the completions endpoint I get 404: {
"error": "Model with requested name is not found"
}

Step-by-step reproduction

No response

Relevant log output

No response

Issue submission checklist

• I'm reporting an issue. It's not a question.
• I checked the problem with the documentation, FAQ, open issues, Stack Overflow, etc., and have not found a solution.
• There is reproducer code and related data files such as images, videos, models, etc.

[Iffa-Intel]

@fedecompa I encountered several issues too when attempting the steps in this guide (which you shared) on Windows: How to serve LLM models with Continuous Batching via OpenAI API.

Please note that this demo was officially validated on Intel® Xeon® processors Gen4 and Gen5 and Intel dGPU ARC and Flex models on Ubuntu22/24 and RedHat8/9. Other OS/hardware might work but still, issues are expected.

[fedecompa]

@Iffa-Intel thanks for the reply.
Actually the GPU is detected correctly from the docker container running on the WSL2 Ubuntu22.
And also the model is running correctly with the OVModelForCausalLM library for python on windows locally:

model_id = "Fede90/llama-3.2-3b-instruct-INT4"
model = OVModelForCausalLM.from_pretrained(model_id, device="GPU.0", trust_remote_code=True)

So it is actually very strange...

[Iffa-Intel]

@fedecompa we'll further investigate & clarify this and get back to you. This probably relates to the architecture of WSL2 in Windows vs Ubuntu which influenced the OpenVINO library functionality.