vllm

Add `--max-model-len auto` to auto-fit context to available memory

#29431

Merged