Run the Custom Image on the Inductiva API ⚡️

With your custom container uploaded to Inductiva, you're now ready to run simulations — no more waiting for model downloads at runtime! 🎉

Using the Image in a Simulation

You can specify your uploaded container in an inductiva.simulators.CustomImage object. Here's an example of how to define and run a task using your diffdock-with-models image:

import inductiva

# Define the machine group
mg = inductiva.resources.MachineGroup(
    machine_type="g2-standard-8",
    num_machines=1,
    spot=True,
)

# Define the simulator using your custom image
diffdock_simulator = inductiva.simulators.CustomImage(
    container_image="inductiva://my-containers/diffdock-with-models"
)

# Define input parameters and run the task
task = diffdock_simulator.run(
    input_dir="path/to/input",
    on=mg,
    commands=[
        'micromamba run -n diffdock python /home/appuser/DiffDock/inference.py '
        '--config ./custom_model_path_inference_args.yaml '
        '--protein_sequence GIQSYCTPPYSVLQDPPQPVV '
        '--ligand "COc(cc1)ccc1C#N" '
        '--samples_per_complex 5'
    ]
)

💡 Tip: Ensure that any necessary config files or inputs (like custom_model_path_inference_args.yaml) are included in your input directory.

Experiments on samples per complex

In order to explore the effect of varying the samples_per_complex parameter on runtime and cost, we conducted a series of benchmark tests using two machine types: a2-highhpu-1g (NVIDIA A100) and g2-standard-8 (NVIDIA L4). This parameter controls the number of binding poses generated per protein-ligand pair, which directly impacts computational load.

The following table summarizes the duration and cost for processing different numbers of samples on each machine type:

These results indicate that while a2-highhpu-1g achieves marginally faster runtimes, the g2-standard-8 offers substantially better cost-efficiency, especially at larger sample sizes. Therefore, for many use cases, g2-standard-8 may be the preferred option when optimizing for price-performance trade-offs.

Wrapping Up

With this workflow, you've created a fast, portable, and scalable version of DiffDock for use with Inductiva. By preloading models in a custom Docker image and uploading it to your personal container registry, you've eliminated costly download steps during inference, resulting in faster and more reliable cloud simulations.

You're now equipped to run custom deep learning workloads with maximum efficiency on Inductiva! ✅