Skip to main content
boileroom supports two backends for executing models on GPUs. The user-facing API is identical regardless of backend — you call .fold() or .embed() the same way, and get the same output types back.

​
Backend selection

Set the backend when creating a model: 
from boileroom import ESMFold

# Serverless GPU via Modal (default)
model = ESMFold(backend="modal")

# Local GPU via Apptainer container
model = ESMFold(backend="apptainer")
Modal provides serverless GPU execution. Your code runs locally, but the model inference runs on Modal’s cloud GPUs. Containers scale to zero when idle and spin up on demand.

​
Setup

  1. Install Modal: pip install modal
  2. Authenticate: modal token new
  3. Use the model with backend="modal" (this is the default)

​
GPU options

Modal automatically selects a GPU based on the model’s requirements. The following GPU types are available on Modal:
GPUVRAM
T416 GB
L424 GB
A10G24 GB
A100-40GB40 GB
A100-80GB80 GB
L40S48 GB
H10080 GB

​
Serverless scaling

  • Containers cold-start when first called (may take 30-60 seconds)
  • Subsequent calls reuse warm containers
  • Containers automatically scale down after a period of inactivity

​
Apptainer backend

Apptainer (formerly Singularity) runs models in a container on your local machine or HPC cluster. This is useful when you need local execution, cannot use cloud services, or are running on institutional compute.

​
Setup

  1. Install Apptainer on your system
  2. Use backend="apptainer" or backend="apptainer:tag" to specify a Docker image tag

​
Tag syntax

By default, the Apptainer backend uses the latest Docker image tag. You can specify a different tag: 
# Use the latest image
model = ESMFold(backend="apptainer")

# Equivalent to above
model = ESMFold(backend="apptainer:latest")

# Use a specific version tag
model = ESMFold(backend="apptainer:v1.0.0")

​
Local execution

The Apptainer backend runs the model on your local GPU. Ensure you have a CUDA-compatible GPU with sufficient VRAM for the model.

​
Context manager

Both backends support context manager usage to ensure proper cleanup: 
with ESMFold(backend="modal") as model:
    result = model.fold("MKTVRQERLKSIVRI")
# Backend resources are automatically released
Without a context manager, the backend is cleaned up when the ModelWrapper instance is garbage collected or when the Python process exits.

​
Device selection

The device parameter works the same way across backends: 
# Let the backend auto-select (default: "cuda:0" if available)
model = ESMFold(backend="modal")

# Specify a device
model = ESMFold(backend="modal", device="cuda:0")

# Force CPU execution
model = ESMFold(backend="modal", device="cpu")