vLLM Supports GGUF Models via Plugin, but Only on GPU Hardware

The vLLM inference engine can run GGUF-format models, but only through a separate official plugin called vllm-gguf-plugin and exclusively on supported GPU hardware, including NVIDIA and AMD cards. GGUF support was removed from vLLM's core package and migrated to this out-of-tree plugin, meaning a standard pip install of vLLM is no longer sufficient. Users must also specify a tokenizer flag pointing to the base model, as GGUF tokenizer conversion is documented as slow and unreliable, particularly for models with large vocabularies. Unlike llama.cpp, which dequantizes models on CPU, vLLM's GGUF path relies entirely on GPU kernels designed for batched serving, making CPU inference unsupported. The official vLLM documentation flags GGUF support as highly experimental and under-optimized, with narrower quantization coverage and slower adoption of new model architectures compared to llama.cpp.