Every Reactor model must be decorated with @model. This decorator is metadata: it tells
Reactor what your model is and how to identify it. It does not contain runtime arguments or
configuration values for your model logic.
from reactor_runtime import VideoModel, model
@model(name="my-video-model")
class MyVideoModel(VideoModel):
def start_session(self):
...
Required Fields
name
The only required field. This is the unique identifier for your model in the Reactor ecosystem.
@model(name="my-diffusion-model")
class MyVideoModel(VideoModel):
...
name:
- Use lowercase letters, numbers, and hyphens
- No spaces or special characters
- Should be unique across your models
- This is used in logs, deployments, and URLs
- This is what clients will use to identify your model and connect to it
Examples:
@model(name="longlive")
@model(name="my-diffusion-v2")
@model(name="world-simulator-beta")
Optional Fields
config (Default Configuration File)
@model(name="my-model", config="configs/default.yaml")
class MyVideoModel(VideoModel):
...
__init__.
This is the factory configuration that ships with your model when it is built and deployed.
If you do not specify config, your model receives an empty configuration (unless the user passes
-c on the command line).
Configuration values can be overridden via CLI flags like --model.fps=60. See
Configuration for details on overrides and best practices. weights
@model(name="my-model", weights=["my-model-weights", "vae-decoder"])
class MyVideoModel(VideoModel):
...
weights field
becomes important when deploying to Reactor’s infrastructure.
You can pass any additional keyword arguments to store custom metadata:
@model(
name="my-model",
config="configs/default.yaml",
author="Your Name",
version="1.0.0",
description="A real-time diffusion model"
)
class MyVideoModel(VideoModel):
...
Running Your Model
When you run reactor run, the CLI scans Python files in the current directory for a class
decorated with @model. Navigate to the folder containing your model file and run:
cd my-model
reactor run --runtime http
- Import Python files in the directory
- Find the class decorated with
@model
- Read the decorator’s configuration
- Instantiate your model and start the runtime
Only one @model decorator per project. If multiple classes are decorated with @model in
the same directory, Reactor will raise an error. Each project should have exactly one entry point.
Import Errors
If your model file has import errors before the decorated class definition, the CLI will report
that no decorated class was found. This happens because Python cannot parse the file far enough to
see the decorator.
Error: No @model decorated class found in project.
- Missing dependencies (
pip install your requirements)
- Syntax errors earlier in the file
- Import errors in modules your file depends on
Fix the import errors first, then reactor run will find your decorated class.
Complete Example
from reactor_runtime import VideoModel, model, command, get_ctx
from omegaconf import DictConfig
@model(
name="realtime-diffusion",
config="configs/production.yaml",
weights=["diffusion-weights", "vae-decoder"]
)
class RealtimeDiffusionModel(VideoModel):
def __init__(self, config: DictConfig):
self.num_steps = config.get("num_steps", 5)
self.pipeline = self._load_pipeline()
def _load_pipeline(self):
weights_path = VideoModel.weights("diffusion-weights")
# Load your model...
@command("set_prompt", description="Set the generation prompt")
def set_prompt(self, prompt: str):
self.current_prompt = prompt
def start_session(self):
while not get_ctx().should_stop():
frames = self.generate()
get_ctx().emit_block(frames)
Decorator vs Configuration
It is important to understand what goes where:
@model Decorator | Configuration (YAML file) |
|---|
| What your model is | How your model runs |
| Model identity and metadata | Runtime parameters |
| Weight folder names | FPS, resolution, steps |
| Does not change per environment | Can vary per deployment |
@model(name="my-diffusion", weights=["my-weights"])
fps: 30
num_steps: 8
guidance_scale: 7.5
