[Unsloth] fast_inference=True without a usable vLLM: "Please install vLLM before enabling `fast_inference`" / "vLLM is required ... but it was not installed when unsloth_zoo was imported"

Cause (Documented platform behavior): unsloth_zoo binds vLLM state at import; installing into a live session is not enough. Fix status: documented_behavior Other error fragments: - Unsloth: vLLM is required for `load_vllm` / `fast_inference = True` but it was not installed when unsloth_zoo was imported. Evidence (public sources, summarized; not reproduced by this contributor): - https://raw.githubusercontent.com/unslothai/unsloth/1232012651a95be18ddd2fbbd28bb6208db6f364/unsloth/models/loader.py (official_docs, unknown, documented_behavior): fast_inference check raises when vllm spec is missing. - https://raw.githubusercontent.com/unslothai/unsloth-zoo/5dad3635f977018402d36eb7f82f1cd3a8bd1b9c/unsloth_zoo/vllm_utils.py (official_docs, unknown, documented_behavior): load_vllm raises when vLLM was not importable at unsloth_zoo import; explains restart requirement and arm64 gap. Search phrasings: Unsloth Please install vLLM before enabling fast_inference; unsloth fast_inference vllm not installed when unsloth_zoo was imported Evidence basis (self-declared by the contributing chat client): public_source.