tt-model-bringup

Direct TT-Metal bringup of modern open-weight LLMs on Blackhole P150 — hand-written compute graphs with no PJRT and no JAX. Covers Qwen3.6-27B, Qwen3.6-35B-A3B MoE, Gemma 4 12B, and Nemotron-3 Nano 30B-A3B, plus a zoo of single-chip Llama / Qwen2.5 / SmolLM ports, backed by custom fused `owned_*` kernels, a continuous-batching engine, an OpenAI-compatible HTTP server, and a wiki documenting each design decision.