| # Setting up your local ColabFold API server |
|
|
| Here you will find two examples of how to setup your own API server on a Linux (or macOS for testing) machine. |
|
|
| ## `setup-and-start-local.sh` |
|
|
| The `setup-and-start-local.sh` script will execute most of the steps to get a server running for you. |
| It will do the following steps: |
| * check that all required software is installed (`curl`, `aria2c`, `rsync`, `aws`) |
| * download pinned **MMseqs2** and **mmseqs-server** binaries for your platform (Linux x86\_64/arm64, macOS universal) |
| * download the databases (UniRef30 and ColabFoldDB, this might take some time) |
| * download the API server and compile its binary |
| * start the API server |
| |
| The script can be called repeatedly to (re)start the server. It avoids unnecessary work and only re-downloads components when the pinned commit changed. |
| |
| ### CPU/GPU and platform detection |
| |
| * Uncomment the `export GPU=1` line to enable GPU mode (Linux only). |
| * The script adds the parameters `--paths.colabfold.gpu.gpu 1 --paths.colabfold.gpu.server 1`. See `config.json` for more details. |
| |
| ### Choosing a PDB rsync mirror |
| |
| At the top of the script you can set the PDB mirror to use (RCSB, PDBe or PDBj). |
| Uncomment the pair you want. The script exits if no mirror is selected. |
| |
| ### Configuration |
| |
| Edit `config.json` as needed. Common tweaks: |
| |
| * `server.address` — change the bind address/port (we recommend putting `nginx` in front for gzip/SSL). |
| * `local.workers` — number of local job workers. |
| * Optional GPU block under `paths.colabfold.gpu` lets you pin device IDs per DB when you run multi-GPU. |
| * A `server.ratelimit` example is included and can be enabled. |
| |
| ### Run |
| |
| ``` |
| ./setup-and-start-local.sh |
| ``` |
| |
| If `DEBUG_MINI_DB=1` is set, the server starts with templates disabled and a tiny DB for quick tests. |
| |
| ## Setup a systemd service |
| To better manage the ColabFold API server, we recommend to setup a systemd service. It will automatically restart on failure and lets you use `journalctl`/`systemctl`. |
| |
| 1. First run `setup-and-start-local.sh` once to get the folder structure and binaries. |
| 2. Adjust the `systemd-example-mmseqs-server.service` example and point it to your paths: |
| 3. Enable and start `./restart-systemd.sh` |
| |
| ## Forcing databases to stay resident in system memory |
| |
| The ColabFold MSA API server will only achieve response time of few seconds if the search database are held fully within system memory. We use vmtouch (https://github.com/hoytech/vmtouch) to keep the precomputed database index file within system memory. In CPU mode, this is the most expensive part of the MSA API server, as the two default databases (UniRef30+ColabFoldDB) require currently 768GB-1024GB RAM to stay resident in RAM and have enough RAM spare for worker processes. |
| |
| After installing `vmtouch`, you can execute the following command to make sure that the search databases are not evicted from the system cache: |
| |
| ``` |
| cd databases |
| sudo vmtouch -f -w -t -l -d -m 1000G *.idx |
| ``` |
| |
| ## Using a custom API server |
| |
| You can pass the server URL to `colabfold_batch` via `--host-url`. |
| In notebooks, add `host_url=https://yourserver.example.org` to the `run()` call in the *Run Prediction* cell. |
|
|
