tt-tinygrad

A Tenstorrent backend for tinygrad that targets TT-Lang rather than raw tt-metal: a Renderer classifies each UOp kernel graph as matmul, reduce, or elementwise and emits ttl.math.* Python source, and a Compiled device parses the rendered kernel's contract, materializes ttnn tensors from host buffers, and calls it in-process. Proof of concept — 110 pass / 13 xfail across 125 cases on a QuietBox, covering fused matmul, reductions, softmax, layernorm, and attention chains, on top of a three-line patch to upstream tinygrad.