BioNeMo Recipes
Type: Reference implementation Tags: NVIDIA, BioNeMo, biological foundation models, drug discovery, training recipes, Transformer Engine, PyTorch, FSDP, FP8, life sciences Related: NVIDIA-BioNeMo, NVIDIA-Clara, NVIDIA-AI-Enterprise, NGC, Transformer-Engine, PyTorch, Hugging-Face-Accelerate, Megatron-LM, NVIDIA-NeMo, NIM-for-Evo-2, NIM-for-ProteinMPNN, NIM-for-RFdiffusion, NIM-for-DiffDock Sources: https://docs.nvidia.com/bionemo-framework/latest/main/recipes/, https://docs.nvidia.com/bionemo-framework/latest/main/recipes/recipes/, https://docs.nvidia.com/bionemo-framework/latest/models/index.html, https://docs.nvidia.com/bionemo-framework/latest/models/amplify/, https://docs.nvidia.com/bionemo-framework/latest/models/ESM-2/, https://docs.nvidia.com/bionemo-framework/latest/models/ESM-2/pre-training/, https://docs.nvidia.com/bionemo-framework/latest/models/evo2/, https://docs.nvidia.com/bionemo-framework/latest/models/geneformer/, https://docs.nvidia.com/bionemo-framework/latest/main/recipes/models/amplify/amplify/, https://docs.nvidia.com/bionemo-framework/latest/main/recipes/models/esm2/esm2/, https://docs.nvidia.com/bionemo-framework/latest/main/recipes/models/geneformer/geneformer/ Last Updated: 2026-04-29
Summary
BioNeMo Recipes is NVIDIA’s public reference-implementation layer for scaling biological foundation model training on NVIDIA GPUs. Current docs position it beside the BioNeMo Framework: instead of being a broad batteries-included training framework, the recipes provide optimized model checkpoints, Transformer-Engine-based model implementations, and self-contained PyTorch training examples that researchers can adapt for biological AI workloads.
Detail
Current positioning
The latest BioNeMo docs note that the documentation is being refactored while NVIDIA consolidates earlier 5D parallelism training material with bionemo-recipes. For wiki purposes, this page is the canonical place for the durable recipe layer rather than separate pages for every recipe folder, model variant, notebook, or helper script.
NVIDIA-BioNeMo remains the platform page for the broader life-sciences ecosystem, including BioNeMo Framework and BioNeMo NIMs. This page focuses on the training and reference-code layer.
What the recipes provide
- Hugging Face-compatible
PreTrainedModelclasses that use NVIDIA Transformer-Engine layers internally. - Self-contained Docker recipe directories with training scripts, pinned dependencies, test contracts, and example multi-node launch material.
- Recipe coverage for PyTorch-native training loops, Hugging-Face-Accelerate, PyTorch Lightning, FSDP2, and megatron-FSDP style scaling.
- Training-feature examples around BF16, FP8, MXFP8, sequence packing/THD input format, context parallelism, checkpoint conversion, and Hugging Face checkpoint import/export.
- Performance-oriented examples and benchmarking expectations across NVIDIA data center GPUs such as A100, H100, H200, B100, and B200.
Current model and recipe scope
Current BioNeMo Framework model docs list AMPLIFY, ESM-2, Evo2, and Geneformer. The recipe docs additionally call out optimized AMPLIFY, ESM-2, Geneformer, CodonFM, and vision-transformer examples.
Important boundary: several biological models surfaced here come from third-party research communities and are not NVIDIA-owned model families. The wiki keeps those model facts folded into this BioNeMo recipe page unless NVIDIA’s docs make a durable NVIDIA service, NIM, or product page for that topic. For example, NIM-for-Evo-2, NIM-for-ProteinMPNN, NIM-for-RFdiffusion, and NIM-for-DiffDock remain separate because NVIDIA publishes them as NIM services.
Developer workflow
BioNeMo Recipes separates reusable models from runnable recipes:
models/packages are intended to be loadable through standard Hugging Face model APIs after conversion or publication.recipes/folders are educational and self-contained, with Dockerfiles, Hydra-style configuration, training entrypoints, small validation data where appropriate, tests, and optional SLURM examples.- CI/CD expectations are intentionally simple: build the component container and run its pytest suite.
- The docs emphasize readable, adaptable code over a highly abstracted shared framework, so researchers can modify the full training loop for their own biological datasets.
Connections
- NVIDIA-BioNeMo - broader platform page for BioNeMo Framework, BioNeMo NIMs, and life-sciences AI workflows.
- NVIDIA-Clara - healthcare and life-sciences umbrella that includes BioNeMo, Parabricks, MONAI, and medical imaging AI.
- NVIDIA-AI-Enterprise - enterprise support and production software context for BioNeMo and NIM deployment.
- NGC - distribution channel for NVIDIA containers, model assets, and private-registry workflows used around BioNeMo.
- Transformer-Engine - NVIDIA library providing TE layers, FP8/MXFP8/NVFP4 recipes, and optimized transformer building blocks used by BioNeMo Recipes.
- PyTorch - primary deep learning framework underneath the recipe training loops.
- Hugging-Face-Accelerate - one of the recipe execution paths for distributed PyTorch training.
- Megatron-LM - adjacent NVIDIA scaling lineage for large transformer training and megatron-FSDP style patterns.
- NVIDIA-NeMo - related NVIDIA training and model-lifecycle ecosystem; BioNeMo adopts similar large-scale training concepts for biological models.
- NIM-for-Evo-2, NIM-for-ProteinMPNN, NIM-for-RFdiffusion, and NIM-for-DiffDock - inference-service pages for durable BioNeMo/NIM topics that are separate from recipe code.
Source Excerpts
- NVIDIA docs describe BioNeMo Recipes as reference implementations for scaling biological foundation model training with Transformer Engine and FSDP.
- Current BioNeMo Framework model docs list AMPLIFY, ESM-2, Evo2, and Geneformer as available models.
- Recipe docs distinguish model packages from self-contained training recipes and document Docker/test/benchmark expectations.