Skip to main content

Modal Integration

Run Modal functions, apps, and serverless GPU workloads from Fused UDFs. Your Modal credentials are stored securely as Fused secrets — no tokens need to be hardcoded.

Prerequisites​

  1. A Fused account on a paid plan (execution environment required).
  2. A Modal token. Generate one at modal.com/settings/tokens or run modal token new locally — you will get a Token ID (starts with ak-) and a Token Secret (starts with as-).

Setup​

Add your Modal tokens​

Open Settings > Integrations & Secrets, find the Modal section, and paste both values:

  • MODAL_TOKEN_ID
  • MODAL_TOKEN_SECRET

Each is stored as a Fused secret under the corresponding name.

Install the SDK (optional)​

The modal package is pre-installed in the Fused runtime. If you are working locally, install it with:

pip install modal

Quick start​

fused.api.modal_connect() returns an authenticated modal.Client, pre-configured with the credentials stored in fused.secrets. From there you can look up deployed Modal apps and functions and invoke them.

@fused.udf()
def udf():
    import fused
    import modal

    # Authenticate using the stored Modal token
    fused.api.modal_connect()

    # Look up a deployed function and call it
    embed = modal.Function.from_name("my-app", "embed_text")
    return embed.remote("Hello from Fused!")

API reference​

fused.api.modal_connect() -> modal.Client​

Return a modal.Client initialized from fused.secrets["MODAL_TOKEN_ID"] and fused.secrets["MODAL_TOKEN_SECRET"].

import fused
client = fused.api.modal_connect()

Calling modal_connect() configures the Modal SDK for the current process — subsequent calls to modal.Function.from_name(...).remote(...), modal.Cls.from_name(...), etc. authenticate automatically.

See the Modal Python SDK reference for the full API surface.

Common operations​

Invoke a deployed function​

@fused.udf()
def udf(prompt: str = "A high-resolution satellite image of a coastline"):
    import fused
    import modal

    fused.api.modal_connect()

    generate = modal.Function.from_name("image-gen", "generate")
    image_bytes = generate.remote(prompt)
    return image_bytes

Map over inputs in parallel​

Use .map() to fan out work across Modal's serverless GPUs.

@fused.udf()
def udf():
    import fused
    import modal
    import pandas as pd

    fused.api.modal_connect()

    embed = modal.Function.from_name("embeddings", "embed")
    texts = ["hello", "world", "fused", "modal"]
    vectors = list(embed.map(texts))

    return pd.DataFrame({"text": texts, "embedding": vectors})

Call a class-based service​

@fused.udf()
def udf():
    import fused
    import modal

    fused.api.modal_connect()

    Model = modal.Cls.from_name("llm-app", "Llama")
    model = Model()
    return model.complete.remote("Explain GeoParquet in one paragraph.")

Asynchronous invocation​

For long-running jobs, spawn the call and poll for the result instead of blocking the UDF.

@fused.udf()
def udf():
    import fused
    import modal

    fused.api.modal_connect()

    train = modal.Function.from_name("trainer", "train_model")
    call = train.spawn(epochs=10)
    return {"call_id": call.object_id}

You can then retrieve the result in a follow-up UDF with modal.FunctionCall.from_id(call_id).get().

Disconnecting​

To revoke the integration, go to Settings > Integrations & Secrets, find the Modal section, and remove your tokens. You should also revoke the token in the Modal dashboard if it is no longer needed.