stream.drivers.gpt

stream.drivers.gpt

OpenAI-compatible (GPT / vLLM / local) driver -- portable, no Cursor dependency.

Uses the Chat Completions tool-calling loop. base_url lets you point at any OpenAI-compatible endpoint (e.g. a local vLLM server hosting an open-weight model), which is also the reference setup for rich logprobs: score requests logprobs and returns the top token with its probability.