hao-ai-lab-FastVideo/docs/inference/configuration.md

Configuration

Multi-GPU Setup

FastVideo automatically distributes the generation process when multiple GPUs are specified:

# Will use 4 GPUs in parallel for faster generation
generator = VideoGenerator.from_pretrained(
    "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
    num_gpus=4,
)

One node uses the multiprocessing executor (execution_backend: mp, the default). With num_gpus=1, mp automatically selects the in-process (uni) executor so weights load in the current process. Set execution_backend: uni to force that path. Two machines — for example two DGX Sparks, one GPU each — need Ray:

generator:
  engine:
    num_gpus: 2
    execution_backend: ray
    parallelism:
      sp_size: 2

Set RAY_ADDRESS and FASTVIDEO_HOST_IP to the interconnect IPs, not Wi-Fi. The FastH3 example selects Ray automatically when RAY_ADDRESS is set. Full bring-up: Pair two NVIDIA DGX Sparks.

Customizing Generation

• PipelineConfig: Initialization time parameters
• SamplingParam: Generation time parameters

You can customize generation behavior using PipelineConfig and SamplingParam:

from fastvideo import VideoGenerator, SamplingParam, PipelineConfig

def main():
    model_name = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
    config = PipelineConfig.from_pretrained(model_name)
    config.vae_precision = "fp16"

    # Create the generator
    generator = VideoGenerator.from_pretrained(
        model_name,
        num_gpus=1,
        dit_layerwise_offload=True,  # FastVideoArgs option
        pipeline_config=config
    )

    # Create and customize sampling parameters
    sampling_param = SamplingParam.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers")
    
    # How many frames to generate
    sampling_param.num_frames = 45
    
    # Video resolution (width, height)
    sampling_param.width = 1024
    sampling_param.height = 576
    
    # How many steps we denoise the video (higher = better quality, slower generation)
    sampling_param.num_inference_steps = 30
    
    # How strongly the video conforms to the prompt (higher = more faithful to prompt)
    sampling_param.guidance_scale = 7.5
    
    # Random seed for reproducibility
    sampling_param.seed = 42  # Optional, leave unset for random results

    # Generate video with custom parameters
    prompt = "A beautiful sunset over a calm ocean, with gentle waves."
    video = generator.generate_video(
        prompt, 
        sampling_param=sampling_param, 
        output_path="my_videos/",  # Controls where videos are saved
        save_video=True
    )

    # If return_frames=True, frames are available in video["frames"]
    print(f"Generated {len(video['frames'])} frames")

if __name__ == '__main__':
    main()

JSON/YAML Config Files (CLI)

The inference CLI is config-first. Use an explicit subcommand with --config, then apply optional dotted overrides on top, matching the training CLI style. By default, CLI generation uses return_frames=false unless you set request.output.return_frames: true in config or via a dotted override.

fastvideo generate --config config.yaml

Example nested config:

generator:
  model_path: FastVideo/FastHunyuan-diffusers
  engine:
    num_gpus: 2
    parallelism:
      sp_size: 2
request:
  prompt: A capybara relaxing in a hammock
  sampling:
    num_frames: 45
    height: 720
    width: 1280
    num_inference_steps: 6
    seed: 1024
  output:
    output_path: outputs/

generator.engine.execution_backend accepts mp, uni, ray, or external_launcher. The last option is for synchronized offline generation under torchrun or srun; it is not supported by fastvideo serve. Set generator.engine.num_gpus and the relevant parallelism sizes to the total launcher world size. See the distributed inference contract and the MiniMax-H3 run config.

generator.engine.parallelism.dist_timeout is an optional positive integer in seconds. FastVideo applies it to the default process group when FastVideo initializes that group, and to every FastVideo-created device and Gloo control group. If it is omitted, PyTorch's backend defaults apply. This timeout bounds a missing collective participant; external launchers should still supervise the full worker group and terminate peers when one rank exits.

Override individual values from the CLI with dotted paths:

fastvideo generate --config config.yaml --request.sampling.seed 42

Performance Optimization

For configuring optimizations, please see our optimizations guide