moritztng

Playbook for implementing/porting a new biomolecular model (protein folding, structure prediction, protein language model, diffusion structure head, MSA encoder, etc.) onto Tenstorrent hardware inside the tt-bio codebase. Use when the task is "port model X to ttnn/Tenstorrent", "add a new model to tt-bio", "implement <bio model> on TT", or extending an existing port. Encodes the non-negotiable requirements (inference-only, accuracy, performance, --fast, unification, no redundancy, vendored pip deps) and the hard-won methodology (component-by-component parity porting, on-device residency, warm profiling, bit-exact bucketing, sample-dim batching with OOM chunking, bf16 load). Distilled from the ESMC-300M/6B + ESMFold2 and Boltz-2 ports (the 58-commit ESMFold2 history).

Predict 3D biomolecular structures and binding affinity (Boltz-2, ESMFold2, Protenix, OpenDDE) and design de-novo binders/proteins (BoltzGen) via JapanFold — a free, public, Tenstorrent-accelerated HTTP API. Use to fold a protein or complex, co-fold a protein with a ligand and get affinity, fold an antibody-antigen complex, design nanobody/antibody/peptide/miniprotein binders against a target, turn a sequence into a PDB/mmCIF structure, or compute ESMC protein embeddings (per-residue + pooled vectors). No API key or local GPU needed.