Native-speed vLLM transformers modeling backend

vLLM says its upgraded transformers backend can now match or beat many hand-written native vLLM implementations on several Qwen3 setups.

A single flag lets Hugging Face models run in vLLM while still using its parallelism and inference optimizations.

The result lowers the need for custom vLLM model ports and makes high-performance serving easier for more architectures.