Blog

Notes on the UniLLM internals — the tensor model, the scheduler, the KV cache, and how the 47 architectures collapse into one trait.