Turnkey-LLM (aka lemonade)

[jeremyfowers]

This update adds the turnkey-llm CLI and turnkeyml.llm APIs, which provide LLM-specific tools for prompting, serving, and accuracy tests.

This initial release has device/runtime support for:

• Huggingface LLMs with CPU and CUDA backends
• OnnxRuntime-GenAI with DirectML backend
• RyzenAI-Transformers backend

The tools provided are:

• huggingface-load: load LLM from huggingface checkpoint
• oga-load: load LLM using OnnxRuntime-GenAI onto DirectML device
• ryzenai-npu-load: load LLM onto RyzenAI NPU device
• llm-prompt: chat with your LLM
• serve: launch a websocket server for your LLM
• accuracy-mmlu and accuracy-perplexity: measure the performance of your LLM on the target backend

Learn more about the features and usage in the README.md at src/turnkeyml/llm/README.md.