Support V1 API

[ilya-lavrenov]

See https://github.com/vllm-project/vllm/tree/main/vllm/v1

[popovaan]

V1 currently supports only GPU device for Ampere or later NVIDIA GPUs: https://blog.vllm.ai/2025/01/27/v1-alpha-release.html
CPU is not supported for V1, so we can't support it yet.

[ilya-lavrenov]

I suppose this feature depends on backend.
V1 or V0 is just an API, which can be implemented in each backend independently. E.g. here https://github.com/vllm-project/vllm-ascend/blob/main/vllm_ascend/worker/worker_v1.py we can see worker and model runner from V1 are already implemented. So, it seems we are not blocked from OpenVINO side?

[popovaan]

What I mean is LLMEngine fails on device check before running the worker. But I agree that generally we should be able to override this device check. I will look deeper into this.