huggingface/text-generation-inference ? reverse-engineered prompt

Build me a text generation server that can run popular open source language models like Llama or Mistral from a model name and expose an HTTP API for generating text and streaming tokens back as they are produced.

I want it to feel production ready, so include a simple way to start it with Docker, support for GPU acceleration, and an OpenAI style chat completions endpoint as well as a basic generate endpoint. It should handle multiple requests efficiently, keep responses streaming, and support common model settings like temperature, top p, top k, stop sequences, and max tokens.

Please also add the usual things needed for real deployment, like metrics, tracing, and clear docs for running it locally with a Hugging Face model and token if the model is private. If you need to check current docs for the model or deployment details, go ahead and look them up online.