Add vLLM Engine

[IlyaMescheryakov1402]

Adding vLLM Engine (v 0.7.3) with OpenAI API

Added routes:

• /v1/completions
• /v1/chat/completions
• /v1/models

Entrypoint for vLLM-OpenAI is http://127.0.0.1:8080/serve/openai/v1 (in its basic options). Endpoint name doesn't change entrypoint (as it does for other engines) according to OpenAI API. Endpoint name in that case equals served_model_name that can be send via openai api request.

vLLM engine can be configured via docker/docker-compose-gpu.yml (line 108, variable VLLM_ENGINE_ARGS). Model can be configured via preprocess.py file. For vllm deploy example see examples/vllm/readme.md

Other small fixes:

1. Add CLEARML_DEFAULT_SERVE_SUFFIX - to change entrypoint suffix and link
2. Fix searching control plane task in triton_helper (in case if we have more than 1 serving instance)
3. Add docker-compose-gpu.yml file for instances that serves models on GPU (for custom, custom_async and vllm engines)

[InsertNamePls]

@IlyaMescheryakov1402

image: clearml-serving-inference:latest not being allegroai/clearml-serving-inference:latest will not create login issues?

[IlyaMescheryakov1402]

Hi @InsertNamePls

Thank you for your attention

Yes, it was a mistake. I used custom images built with Dockerfiles during testing, because docker images allegroai/clearml-serving-inference:latest are not suitable for PR testing (they install latest clearml-serving package and don't have vllm as registered engine).

For testing this PR, you have to:

1. comment this string and this string and build your own images (both for serving and for statistics)
2. change images in docker/docker-compose-gpu.yml file (lines 78 and 122 accordingly) on your local ones (that was built on the previous step)
3. run docker-compose file