sampling

Configurable diffusion samplers for RL training methods.

Classes

fastvideo.train.methods.rl.common.sampling.DiffusionSampler

DiffusionSampler(config: SamplingConfig)

Thin model/scheduler sampler used by RL methods.

This intentionally does not call FastVideo's full inference pipelines. RL training needs a reusable sampling primitive that works with ModelBase wrappers and scheduler math without binding a method to model-family pipeline classes such as WanDMDPipeline.

Source code in fastvideo/train/methods/rl/common/sampling.py

def __init__(self, config: SamplingConfig) -> None:
    self.config = config

fastvideo.train.methods.rl.common.sampling.SamplingConfig dataclass

SamplingConfig(num_steps: int = 25, scheduler: SchedulerName = 'model_default', trajectory: TrajectoryName = 'ode', flow_shift: float | None = None, timesteps: list[float] | None = None, sigmas: list[float] | None = None)

YAML-backed sampling knobs shared by RL methods.