Python SDK

The uRun Python SDK ships as urun-cli on PyPI — pip install urun-cli gives you both the urun command (deploy, list, scale) and the importable urun library. It is how you write the backend of a uRun app: ordinary Python functions, decorated with the GPU shape they need, deployed to run on uRun's compute fleet.

The product primitive is a real-time session. You decorate a function with @app.function(...), deploy it, and a browser (via the TypeScript SDK) invokes it by name — getting back a live session with bidirectional media stream(name) and a synced control doc(key). The Python side owns the GPU: it holds a model warm, reads inbound media and control, and produces media out.

The same decorator also covers multi-GPU inference (gpus="h100:8", rank-aware via ctx.rank / ctx.world_size), so one mental model spans single-GPU and sharded real-time apps.

The shape of an app

import urun
from urun import App, Context
from urun.core import Dependencies, Credentials

app = App("my-app")

@app.function(gpus="b200:1", deps=Dependencies(python=["torch"]))
def generate(ctx: Context):
    # long-lived session owner: holds the model warm, reads inbound
    # media + control doc, produces a streamed result out
    ...

Deploy it with the urun CLI, then call it from the browser as app.generate({ prompt: "sunset" }).

Core building blocks

• App & functions — App, the @app.function decorator, and local entrypoints.
• Sessions — session lifecycle: lease, hard caps, idle timeout, scale-to-zero with snapshot wakes, DVR recordings.
• Compute — GPU/CPU specs, warm replicas, and session length (lease / max_session_s).
• Streaming pipelines — ctx.stream, ctx.source | ctx.pipe | ctx.pace | sink, and ctx.doc control.
• Store — content-addressed model/data downloads, versioned artifacts, and @store.cache for keeping models hot.
• Dependencies & Credentials — the remote environment and its secrets.

Get started with the Python SDK