Feature: support biology models - Boltz-2, OpenFold3, ESMFold2

Request type

New model or checkpoint support

Problem and use case

TensorRT Model Connect does not support biomolecular structure prediction. Boltz-2 and OpenFold3 are AlphaFold3-style co-folding systems built around MSA and pair representations, Pairformer-like trunks, and atom-level diffusion. ESMFold2 is architecturally distinct: it conditions folding on ESMC-6B protein-language-model embeddings, with optional MSA conditioning, its own pair trunk, and diffusion-based all-atom generation. They share task-level inputs and outputs, but likely require separate architecture-specific Model Connect families rather than one common graph implementation.

There seems to be TRT backend for Boltz-2 and OpenFold3. However, TensorRT Model Connect does not work out of the box.

• Boltz-2 TensorRT performance reference: https://docs.nvidia.com/nim/bionemo/boltz2/latest/performance.html#table-2-performance-across-optimization-backends
• OpenFold3 TensorRT performance reference: https://docs.nvidia.com/nim/bionemo/openfold3/latest/performance.html#table-2-performance-across-optimization-backends

Proposed outcome

Add a biomolecular structure-prediction task and model-family integration, initially targeting Boltz-2 and OpenFold3.

Acceptance criteria:

• Build a BF16 engine on at least one documented supported NVIDIA GPU covering the complete Pairformer/trunk and diffusion path.
• Support variable sequence lengths, masks, recycling, and documented MSA inputs.
• Return structure and confidence outputs on example input: https://github.com/jwohlwend/boltz/blob/main/examples/prot_custom_msa.yaml
• Validate parity against upstream PyTorch.
• Benchmark warmed steady-state latency against eager PyTorch and torch.compile, excluding engine-build time.

Model and target details

• Boltz-2
• OpenFold3 (gated)
• ESMFold2
• Task: biomolecular structure prediction
• Precision: BF16
• Hardware target: at least one documented supported NVIDIA GPU; additional platform profiles may be qualified separately
• Runtime: TensorRT Model Connect family plugin, bundle, and native runtime

Alternatives considered

No response

Additional context

• Boltz source: https://github.com/jwohlwend/boltz
• OpenFold3 source: https://github.com/aqlaboratory/openfold-3
• ESMFold2 source: https://github.com/Biohub/esm

Submission checks

• I searched open and closed issues and found no duplicate request.
• I removed secrets, private/internal evidence, personal paths, and restricted artifacts.

Start by reviewing the Boltz, OpenFold3, and ESMFold2 source repositories and the linked Boltz example input. Then inspect the existing TensorRT Model Connect family plugin, bundle, and native runtime integration points. Done means BF16 support for the documented GPU, variable lengths, masks, recycling, MSA inputs, structure and confidence outputs, upstream parity, and warmed latency benchmarks.