NVIDIA BioNeMo Recipes Simplify Biology AI Model Training

NVIDIA announced on November 5, 2025, the availability of its BioNeMo Recipes, designed to simplify and accelerate the training of large-scale AI models for biology. These recipes provide step-by-step guides built on PyTorch and Hugging Face, reducing the complexity of training advanced biology-focused AI models.

By leveraging NVIDIA’s Transformer Engine (TE) and Fully Sharded Data Parallel (FSDP), researchers can achieve significant speed and memory efficiency. The recipes integrate these technologies to optimize transformer-style AI models, such as the Hugging Face ESM-2 protein language model, for biological applications.

Key Features of BioNeMo Recipes

Transformer Engine (TE) Integration

The Transformer Engine optimizes transformer computations on NVIDIA GPUs, enabling faster and more efficient model training. It can be seamlessly integrated into existing training pipelines, even for non-standard transformer architectures, by replacing PyTorch modules with TE counterparts and using FP8 autocasting.

FSDP2 Integration

FSDP2 enables auto-parallelism, allowing researchers to scale their models across multiple GPUs without manual intervention. This feature is particularly useful for training large models with billions of parameters, ensuring optimal GPU utilization and performance.

Sequence Packing

Sequence packing improves efficiency by removing padding tokens from input data. By using index vectors to denote boundaries between sequences, this technique reduces memory usage and increases token throughput, enhancing overall training performance.

Real-World Application

Tom Sercu, co-founder and VP of Engineering at EvolutionaryScale, highlighted the critical role of NVIDIA’s Transformer Engine in training ESM3, the largest foundation model for biological data. The integration enabled high throughput and efficient GPU usage, demonstrating the practical benefits of BioNeMo Recipes in real-world applications.

Getting Started with BioNeMo Recipes

To begin using BioNeMo Recipes, users need PyTorch, NVIDIA CUDA 12.8, and the BioNeMo Framework Recipes from GitHub. Detailed documentation is available to guide users through the setup and implementation process. The NVIDIA BioNeMo Collection on the Hugging Face Hub also offers pre-optimized models for immediate use.

Conclusion

NVIDIA’s BioNeMo Recipes represent a significant advancement in simplifying and accelerating the training of large-scale AI models for biology. By integrating cutting-edge technologies like the Transformer Engine, FSDP, and sequence packing, these recipes lower the barrier to entry for researchers and enable more efficient and scalable AI model development.