tiny-llm Turns vLLM Internals Into Coursework

This is a rare course that makes inference performance concrete: you diagnose the bottleneck, implement the fix, and measure the result yourself.

• –Starts with readable Python implementations before moving into C++, Metal, quantization, and optimized kernels
• –Uses Qwen3 and Apple Silicon, making serious inference experimentation accessible without an NVIDIA GPU
• –Covers the serving concepts behind vLLM, including continuous batching, chunked prefill, paged KV caches, and FlashAttention
• –MLX acts as both correctness oracle and performance baseline, forcing students to understand the cost of abstraction
• –Best suited to systems engineers with Python and deep-learning fundamentals; it is coursework, not a production inference server