Indexing Databases for GPU Server

The MSA Search NIM includes GPU-accelerated search through the GPU Server, which requires databases to be indexed with specific parameters. This guide explains how to index custom databases using the MMseqs2 tools included in the NIM container.

When to Index

Databases downloaded from NGC are already pre-indexed and ready for GPU Server use. No additional indexing is required for these databases.

Indexing is only required when:

• You have created a custom MMseqs2 database from FASTA sequences
• You are using third-party databases not downloaded from NGC

Using MMseqs2 From the NIM Container

The NIM container includes MMseqs2 with GPU support. Define a helper function to simplify running MMseqs2 commands:

mmseqs_run() {
  docker run -u $(id -u):$(id -g) --rm --runtime=nvidia --gpus all --entrypoint mmseqs \
    -v "$PWD":/work -w /work nvcr.io/nim/colabfold/msa-search:2 "$@"
}

You can run any MMseqs2 command with that helper function.

For example, to see all available commands:

mmseqs_run --help

Example: Re-Indexing a Downloaded Database

This example demonstrates the indexing process by downloading the PDB70 database, removing its index files, and re-creating them.

1. Install and Authenticate the NGC CLI

The ngc command used below is not included with Docker. Before you run ngc registry model download-version, install the NGC CLI for your platform and register it with your NGC account (for example, ngc config set using your NGC API key). Ensure NGC_API_KEY is exported in your environment if the CLI or download flow requires it. For account setup and Docker registry login, refer to Prerequisites and Support Matrix.

2. Download the Database

DATABASE=pdb70_220313-m18v1
DATABASE_PATH="$PWD/msa-search_v$DATABASE/pdb70_220313"
ngc registry model download-version nim/colabfold/msa-search:$DATABASE
chmod -R 755 "$DATABASE_PATH" # allow to use by other users (for example, docker mounts)
cd "$DATABASE_PATH"

3. Examine the Database Files

ls -la

The pdb70_220313 directory contains:

pdb70_220313/
├── pdb70_220313
├── pdb70_220313.dbtype
├── pdb70_220313.idx
├── pdb70_220313.idx.dbtype
├── pdb70_220313.idx.index
├── pdb70_220313.index
├── pdb70_220313.lookup
├── pdb70_220313_h
├── pdb70_220313_h.dbtype
└── pdb70_220313_h.index

• Core database files: pdb70_220313, pdb70_220313.dbtype, and pdb70_220313_h
• Index files: pdb70_220313.idx and pdb70_220313.idx.index

4. Remove Existing Index Files

rm -f pdb70_220313.idx*

5. Re-Create the Index for GPU Server

mmseqs_run createindex pdb70_220313 tmp --remove-tmp-files 1 --split 1 --index-subset 2
rm -rf tmp

The key parameters for GPU Server indexing are:

• --split 1: Single split for GPU memory
• --index-subset 2: Create GPU-compatible index subset

6. Verify the Index Was Created

ls -la pdb70_220313.idx*

You should see new index files created.

MMseqs2 Command Reference

For comprehensive MMseqs2 documentation, run:

mmseqs_run --help

Or for help on a specific command:

mmseqs_run createindex --help

For additional information, refer to the MMseqs2 User Guide.