GGUFs in transformers natively!

Hugging Face's transformers library now natively supports GGUF (llama.cpp quants) models, allowing users to load them with AutoModelForCausalLM.from_pretrained using a gguf_file parameter. This integration aims to provide smaller, quantized models that fit on laptops, offer PyTorch tooling for debugging, and simplify custom generation. While not replacing llama.cpp for maximum local inference performance, it offers a more flexible environment, with performance on Apple Silicon M2 Max showing competitive token generation rates for models like Qwen3.5-4B Q4_K_M.