Launcher Guardrail to avoid Multiple Awake Instances

[diegocastanibm]

Description:
Currently, the launcher can spawn multiple instances that can remain active simultaneously. This can lead to unpredictable behavior when multiple instances attempt to perform operations on shared resources (GPU's)

Proposed Solution:
Implement a guardrail mechanism to ensure only one launcher instance remains awake at any given time.

cc. @MikeSpreitzer @lionelvillard

[MikeSpreitzer]

This can lead to unpredictable behavior

You mean this can run into memory allocation failures. Different vLLM instances will not overwrite each other's memory.

The launcher's ability to report whether the child process has exited and what it logged should be sufficient for diagnosing memory allocation failures.

Adding a guardrail would require getting the launcher involved in sleeping and waking, which would add complexity and latency.