Se7en Minutes of Silence: What's in the (Prompt) Box? A GPU Would Have Told You Already.
Ornith is an open-source agentic coding model that serves an OpenAI-compatible API, and Claude Code only speaks the Anthropic API, so wiring one to the other means putting a LiteLLM translation proxy in the middle. I built the whole stack on a single EC2 instance with Terraform, first on CPU with llama.cpp and then on GPU with vLLM on an L40S. This post covers the architecture, the gotchas (Claude Code's 43K-token prompt, trust-remote-code, the FlashInfer ninja/nvcc PATH trap, Python 3.11), and a measured CPU-vs-GPU comparison where the same model's prompt processing goes from about 7 minutes to about 4 seconds.
READ ARTICLE