NVIDIA NIM for Boltz-2# Boltz-2 Boltz-2 NIM Overview Advantages of NIMs Release Notes Release 1.0.0 Summary Notes and Limitations Support Matrix Supported Hardware Minimum System Hardware Requirements Supported NVIDIA GPUs Testing Locally Available Hardware Getting Started Prerequisites Installing curl, jq, and the Python requests module NGC Authentication Generate an API key Export the API key Docker Login to NGC Starting the NIM Container Runtime Parameters for the Container Caching Models Stopping the Container Next Steps Inferencing with the NIM Prerequisites Deployment NIM API Reference OpenAPI Specification Example Notebooks Predict a Protein Structure Input Parameters Outputs Examples Basic Protein Structure Prediction Protein-Ligand Complex Prediction Comprehensive Example with All Features Check Readiness Input parameters Outputs Example Configuring the Boltz-2 NIM GPU Selection Environment Variables Volumes Optimization with Boltz2-NIM Automatic Profile Selection Selecting a Profile Manually Enabling or Disabling TensorRT and TensorFloat32 NIM_BOLTZ_PAIRFORMER_BACKEND NIM_BOLTZ_ENABLE_DIFFUSION_TF32 Usage Example Deploying the NIM on a multi-GPU System Adjusting Start-Time NIM Input Limits Environment Variables NIM_MAX_POLYMER_INPUTS NIM_MAX_LIGAND_INPUTS NIM_MAX_POLYMER_LENGTH Usage Optimization Parameters Recycling Steps Sampling Steps Diffusion Samples Step Scale Querying the NIM Repeatedly Running Repeated Queries Against the NIM Serially Performance in Boltz-2 NIM NIM Accuracy Factors Affecting NIM Performance Hardware Factors Input Complexity Model Parameters Performance Characteristics Typical Runtimes Performance Testing Expected Performance Baselines Performance Optimization Tips Troubleshooting Performance Issues Common Issues and Solutions