intel/ipex-llm ? reverse-engineered prompt

Build me a Python library and set of examples that make local LLMs run faster on Intel hardware, especially Intel GPUs, NPU, and CPU.

I want it to work with common model formats and tools people already use, like Hugging Face, llama.cpp, Ollama, vLLM, LangChain, LlamaIndex, and a few finetuning workflows too. The main goal is to let someone take popular models like Llama, Mistral, Qwen, DeepSeek, Phi, Gemma, and similar ones, then run them locally with better speed and lower memory use on Intel machines.

Please include clear install and quickstart docs, simple sample apps, and a few end to end demos for inference, serving, and finetuning. It should feel easy for someone with a Windows or Linux PC to get started, especially on Intel Arc, iGPU, or Core Ultra devices. If you need current setup details from the ecosystem, look them up online and follow the latest docs.