TT-Granite

A TTNN port of IBM's Granite-4.0-H hybrid Mamba2/attention/MoE models (tiny and small) to Wormhole, written as a UCL thesis project and run on submeshes of a Galaxy. Includes two custom TT-Metal kernels, `ssm_update` and `conv1d_decode`, that fuse the Mamba2 decode step, plus decode trace capture and benchmark scripts against HuggingFace CPU/CUDA baselines. With trace enabled the author reports 10.65–10.69 tok/s for tiny on 4 chips (about 18% above an A100) and 5.84–5.89 tok/s for small on 8 chips (about 20% below the A100); all numbers are batch 1, and the fused kernels add only about 1–2%. The thesis PDF documents the porting challenges: quantization, SSM state across chunked prefill, and expert-parallel sharding.