NVIDIA-Digital-Bio

NVIDIA-Digital-Bio / boltz-cp

Public

Context Parallelism Code for Boltz-2

18
1
100% credibility
Found Mar 18, 2026 at 18 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

A proof-of-concept framework enabling distributed training and inference for biomolecular folding models like Boltz using context parallelism across multiple GPUs.

How It Works

1
🔬 Discover Fold-CP

You hear about Fold-CP, a tool from NVIDIA that speeds up predicting 3D shapes of proteins and molecules using multiple powerful computers together.

2
💻 Set up your workstation

Download Fold-CP and get it ready on your multi-GPU computer, so it can share the work across all your graphics cards.

3
🧬 Add your biomolecule data

Upload sequences of proteins, DNA, RNA, or ligands you want to model, and let the tool prepare them automatically.

4
Choose prediction or training
🔮
Predict structures

Generate fast 3D models of your molecules.

🎓
Train a model

Teach the system with your data for even better results.

5
🚀 Launch across GPUs

Hit go and watch the magic as your computers team up to fold proteins lightning-fast, way quicker than before.

✅ Get your results

Enjoy high-quality 3D biomolecule structures or a trained model, perfect for your scientific discoveries.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 18 to 18 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is boltz-cp?

Boltz-cp brings context parallelism to Boltz-2, NVIDIA's biomolecular folding model, enabling distributed training and inference across multiple GPUs via PyTorch's DTensor. It combines 2D context parallelism meshes with data parallelism to handle long sequences—like scalable million-token inference—for protein structures, ligands, and complexes. Users get CLI-driven workflows for torchrun or SLURM jobs, supporting BF16, TF32, and attention backends like cuEquivariance or FlexAttention, without rewriting model code.

Why is it gaining traction?

Unlike sequence or tensor parallelism, context parallelism shards attention along sequence length, pairing well with ring attention patterns and avoiding vLLM-style bottlenecks for bio models. Developers dig the drop-in DTensor integration for Boltz-2, plus a research paper detailing Fold-CP math, making it easy to experiment with GPU clusters (needs 4+ in perfect squares). It's a proof-of-concept upstreaming to official Boltz, so early adopters test PyTorch-native scaling before it hits mainline.

Who should use this?

Bioinformatics engineers training protein-ligand diffusion models on A100/H100 clusters, or ML researchers prototyping context parallelism pytorch for long-context bio tasks. Ideal for teams hitting memory walls on million-residue complexes, needing github workflow integration for CI/CD scaling.

Verdict

Grab it if you're scaling Boltz-2 on multi-GPU setups—solid POC with docs and examples—but at 18 stars and 1.0% credibility, it's experimental; monitor the upstream PR for production stability. (187 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.