Skip to content

modal.md

Latest commit

 

History

History
150 lines (112 loc) · 6.71 KB

modal.md

File metadata and controls

150 lines (112 loc) · 6.71 KB

Modal Serverless Deployment

All serverless scripts live in the modal/ directory and run on Modal with pay-per-second billing.

There are two ways to get started. We strongly recommend Option A — it's free, fast, and identical in MSA quality. Option B exists for legacy/advanced use cases but offers no practical benefit.

Option A: Free MSA Server (Recommended) Option B: Self-Hosted Databases
Setup time ~5 minutes 6-7 hours
Storage cost None ~$120/month (~1.2 TB)
One-time DB cost None ~$30 (download + convert)
MSA speed Same (GPU-accelerated MMseqs2) Same (GPU-accelerated MMseqs2)
MSA quality Identical Identical
Best for Everyone Not recommended

Step 1: Install Modal

pip install modal
modal token new

Step 2: Upload Model Weights (one-time)

modal run modal/upload_weights.py --file /path/to/af3.bin
modal run modal/upload_weights.py --status         # Check weights status

Option A: Use Our Free MSA Server (Recommended)

The fastest way to get started. Our public MSA server handles all sequence database searches for you — no need to download or store the ~1.2 TB of databases yourself. Just upload your model weights and start predicting.

# Single protein
modal run modal/af3_predict.py --input protein.json \
    --msa-server https://romero-lab--alphafold3-msa-server-msa.modal.run

# Batch directory
modal run modal/af3_predict.py --input-dir ./proteins/ --output ./results/ \
    --msa-server https://romero-lab--alphafold3-msa-server-msa.modal.run

# Large batch split across multiple GPUs (each GPU: batch MSA then fold one-by-one)
modal run modal/af3_predict.py --input-dir ./proteins/ --mode multi-gpu --num-gpus 4 \
    --msa-server https://romero-lab--alphafold3-msa-server-msa.modal.run

# Large batch with warm producer-consumer pipeline
modal run modal/af3_predict.py --input-dir ./proteins/ --mode producer-consumer \
    --msa-server https://romero-lab--alphafold3-msa-server-msa.modal.run

That's it — no database setup, no storage costs. The --msa-server flag routes all MSA/template searches to our hosted endpoint, and inference runs on your own Modal GPUs.

Note: The sequence databases (~1.2 TB) are hosted exclusively on our Modal infrastructure and are not available for local download. All MSA/template searches must go through the --msa-server endpoint.

Option B: Self-Hosted Databases (Not Recommended)

Why not? You will spend 6-7 hours downloading and converting ~530 GB of databases, pay ~$30 for the compute, and then ~$120/month to store them on a Modal volume. The MSA results are identical to Option A. There is no speed advantage — both options run the same GPU-accelerated MMseqs2 pipeline on the same databases. Unless you have a very specific reason (e.g., air-gapped environment, custom database modifications), use Option A instead.

1. Prepare Databases (one-time, ~6-7 hours, ~$30)

Downloads ~530GB from Google Cloud and converts to MMseqs2 GPU format.

modal run modal/prepare_databases.py              # Full setup (download + convert)
modal run modal/prepare_databases.py --status      # Check database status
modal run modal/prepare_databases.py --convert-only # Only MMseqs2 conversion

2. Run Predictions (without --msa-server)

# Single protein
modal run modal/af3_predict.py --input protein.json

# Batch directory
modal run modal/af3_predict.py --input-dir ./proteins/ --output ./results/

# Single-GPU mode
modal run modal/af3_predict.py --input-dir ./proteins/ --single-gpu

# Multi-GPU mode (split inputs across 4 GPUs)
modal run modal/af3_predict.py --input-dir ./proteins/ --mode multi-gpu --num-gpus 4

# Warm producer-consumer pipeline
modal run modal/af3_predict.py --input-dir ./proteins/ --mode producer-consumer

CLI Reference

All prediction modes use a single unified script (af3_predict.py):

Flag Default Description
--input - Path to single input JSON file
--input-dir - Path to directory of input JSON files
--output ./af3_output Local output directory
--msa-server - URL of remote MSA server (skips local DB for MSA)
--mode single Pipeline mode: single, multi-gpu, or producer-consumer
--gpu a100 Inference GPU type (l40s, h100, h200, a100)
--num-gpus - Number of GPUs for multi-gpu mode (splits inputs across GPUs)
--producer-gpu l40s Producer GPU type for P-C mode
--num-consumers 4 Number of warm consumer containers (P-C mode)
--parallel 4 Max concurrent GPU invocations (single mode batch)
--batch-size 32 Sequences per MMseqs2 batch
--skip-msa false Skip MSA (input must have pre-computed data)
--msa-only false Only run MSA, skip inference
--single-gpu false Batch MSA + sequential inference on one GPU
--keep-workspace false Don't clean up workspace volume after P-C completion
--check false Check setup status and exit

Examples

# L40S single mode, batch size 128
modal run modal/af3_predict.py --input-dir ./proteins/ --gpu l40s --batch-size 128 \
    --msa-server https://romero-lab--alphafold3-msa-server-msa.modal.run

# H200 single mode, batch size 128
modal run modal/af3_predict.py --input-dir ./proteins/ --gpu h200 --batch-size 128 \
    --msa-server https://romero-lab--alphafold3-msa-server-msa.modal.run

# Multi-GPU with L40S: split across 4 GPUs, batch MSA in groups of 128
modal run modal/af3_predict.py --input-dir ./proteins/ --gpu l40s --batch-size 128 \
    --mode multi-gpu --num-gpus 4 \
    --msa-server https://romero-lab--alphafold3-msa-server-msa.modal.run

# Multi-GPU with H200s: split across 8 GPUs
modal run modal/af3_predict.py --input-dir ./proteins/ --gpu h200 --batch-size 128 \
    --mode multi-gpu --num-gpus 8 \
    --msa-server https://romero-lab--alphafold3-msa-server-msa.modal.run

# L40S producer-consumer, 4 consumers, batch size 128
modal run modal/af3_predict.py --input-dir ./proteins/ --gpu l40s --batch-size 128 \
    --mode producer-consumer --num-consumers 4 \
    --msa-server https://romero-lab--alphafold3-msa-server-msa.modal.run

# H200 producer-consumer, 4 consumers, batch size 128
modal run modal/af3_predict.py --input-dir ./proteins/ --gpu h200 --batch-size 128 \
    --mode producer-consumer --num-consumers 4 \
    --msa-server https://romero-lab--alphafold3-msa-server-msa.modal.run

Cost Estimates

Item Option A (MSA Server) Option B (Self-Hosted)
Database setup Free ~$30 (one-time, 6-7 hours)
Database storage Free ~$120/month (~1.2 TB)
Per prediction (varies) $2-15 $2-15