Key Features

• 3-week structured curriculum covering model implementation, serving infrastructure, and advanced optimizations like paged attention and speculative decoding
• Uses only MLX low-level array APIs — no PyTorch or high-level frameworks — forcing deep understanding of every inference component
• Implements vLLM-style continuous batching, KV cache management, and FP8/INT4 quantization from scratch
• Apple Silicon target leverages Neural Engine and unified memory for accessible yet representative systems-level learning
• Companion book at skyzh.github.io/tiny-llm provides conceptual depth alongside hands-on code