Cache Me If You Can

Hands-on LLM inference tutorial given at ML in PL 2026 in Warsaw by Tenstorrent's Maciek Draguła and Nathan Maurice. Jupyter notebooks run GPT-2 and DeepSeek-R1-Distill-Qwen-1.5B on an n150: prefill versus decode, building a KV cache by hand, and what tt_transformers does under the hood. A setup script installs dependencies and downloads the weights.