Porting from Modal

The SDK exposes a Modal-compatible surface. Most Modal scripts port by changing one import line:

import supercomputer as modal  # was: import modal

app = modal.App("demo")
image = modal.Image.debian_slim().pip_install("torch", "numpy")

@app.function(image=image, gpu="H100", timeout=600)
def train(steps: int) -> float:
    ...

@app.cls(gpu="L4", volumes={"/data": modal.Volume.from_name("weights", create_if_missing=True)})
class Model:
    @modal.enter()
    def load(self): ...
    @modal.method()
    def predict(self, x): ...

@app.local_entrypoint()
def main(steps: int = 100):
    print(train.remote(steps))
    print(Model().predict.remote(1))

Running

Run it the way you would with modal run:

super run demo.py --steps 500        # calls the local_entrypoint
super run demo.py::train --steps 500 # or .remote() a function directly

super run detects a script that declares an App and dispatches it this way; flags after the script are matched to the entrypoint's parameters by name and coerced from its type annotations.

Each (gpu, image, volumes) combination gets one warm session. .remote() pickles the arguments, uploads them, runs the function on the host, streams stdout back live, and returns the unpickled result. A remote exception surfaces locally as supercomputer.RemoteError with the host-side traceback attached. Sessions close when the entrypoint returns.

GPU names

Modal GPU names map to SKUs: "T4", "L4", "A10G", "L40S", "A100", "H100", "H200", "B200" and "RTX-PRO-6000", with an optional count ("A100:8"). gpu=None runs on a CPU-only SKU and "any" picks an L4. An unknown name raises instead of inventing a SKU.

Supported

App, @app.function, @app.cls with @modal.enter / @modal.exit / @modal.method / @modal.parameter, @app.local_entrypoint, .remote() / .spawn() / .map(), the Image builder, Volume.from_name, and Secret.from_dict / from_dotenv / from_local_environ. Options such as cpu=, memory=, retries=, region= and container-scaling knobs are accepted and ignored with a warning.

Not supported

web endpoints, Dict / Queue, schedules, sandboxes, modal deploy, and Secret.from_name (there is no hosted secret store; load secrets from a .env or the local environment instead). .map() runs inputs sequentially on one session; use client.map(...) or super map for a real fan-out (Batch Processing).

Example

A complete runnable port lives in the repo at examples/modal-port/app.py.