[RFC]: Deprecation and removal for `--engine-use-ray`

[simon-mo]

Motivation.

In the async_engine code path, we have an option to launch the engine in a separate process using Ray

        parser.add_argument('--engine-use-ray',
                            action='store_true',
                            help='Use Ray to start the LLM engine in a '
                            'separate process as the server process

Originally, the option make it possible to separate the server's Python overhead with the engine's main scheduler loop.

However, few factors made this unused/less popular

• Ray is an optional component, and typically not used in single node environment.
• The serialization and rpc typically offset the theoretical performance gain
• There are typically other ways to isolate server and engine (through multiprocessing, threading, etc).
• Recently, we are separating this in server using lower overhead approaches [ Frontend ] Multiprocessing for OpenAI Server with zeromq #6883

Proposed Change.

Deprecation of the flag with warning for one release.
Removal of the flag given no major pushbacks.

Feedback Period.

1wk

CC List.

No response

Any Other Things.

No response

[rkooo567]

• 1

[andoorve]

This would greatly simplify some of the logic here:

vllm/vllm/engine/async_llm_engine.py

Line 601 in 6a11fdf

async def run_engine_loop(self):

+1

[wallashss]

Hey, I am looking forward to contributing to this project! I just opened a simple PR that can solve this issue.

Of course, feedbacks are welcome!

Thanks!

[zhuohan123]

+1