---

name: inference description: "Guides FastVideo typed Python and config-first CLI generation, model inputs, outputs, offload, attention backends, quantization, and compilation." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Inference

Use for text-to-video, image-to-video, video-to-video, image, and supported audio generation. Start with a registered model ID and a backend-compatible configuration.

Preferred typed path

from fastvideo import VideoGenerator

generator = VideoGenerator.from_pretrained(
    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", num_gpus=1
)
result = generator.generate({
    "prompt": "A fox running through snow",
    "sampling": {"num_frames": 81, "height": 480, "width": 832, "seed": 42},
    "output": {"output_path": "outputs/", "save_video": True},
})

For a reusable file, use VideoGenerator.from_file() or the CLI. Read API reference, workflows, and optimizations before changing advanced settings. Run typed helper only after choosing a model and accepting remote-weight execution.

CLI

fastvideo generate --config run.yaml
fastvideo generate --config run.yaml --request.sampling.seed 7

The config must be nested; flat flags are not the current contract. Input images, video paths, prompt files, control tensors, refinement inputs, and output modes belong in request.inputs, request.sampling, and request.output.

Performance and safety

Fix prompt, seed, dimensions, steps, and model before comparing backends. Warm up once before measuring torch.compile. Use offload or multi-GPU only when the selected model supports it. Attention and low-bit paths can alter numerics; validate quality on the target configuration rather than assuming equivalence.

See inference troubleshooting for shape, backend, memory, output, and deprecated-API failures. Serving is a separate route: serving.