port-bio-model-to-tenstorrent

Porting protein folding and structure prediction models to Tenstorrent hardware uses a phase-driven workflow within the tt-bio stack, enforcing inference-only code, vendored dependencies, and parity testing for unified device integration.

On-device residency for biomolecular structure prediction ensures that the entire inference workflow of ported architectures runs natively on Tenstorrent hardware. This is a non-negotiable requirement enforced during the tt-bio integration process.

No, you cannot use training code when porting protein folding models to Tenstorrent. The tt-bio stack enforces a strict zero training code requirement, allowing exclusively inference-only implementations for biomolecular architectures.

Yes, you must vendor dependencies for ESMFold2 inspired architectures on Tenstorrent. The tt-bio stack mandates vendored dependencies to minimize technical debt and ensure reproducible inference-only porting.

Parity testing for LM-based components ported to Tenstorrent validates that the hardware implementation matches reference model outputs. This testing phase is mandated within the tt-bio stack to achieve unified integration.

Limitations when porting Boltz-2 architectures include a strict requirement for zero training code and mandatory bf16 or fast mode support. The porting process enforces these constraints to maintain a single shared integration layer.