Custom num_frames during inference for Wan2.2

[1plus1equal3]

Currently, changing num_frames (or any generation-related config) requires re-initializing the runner via create_generator, which reloads the entire model weights:

def create_generator(
    self,
    attn_mode="flash_attn2",
    infer_steps=50,
    num_frames=81,
    height=480,
    width=832,
    guidance_scale=5.0,
    sample_shift=5.0,
    fps=16,
    aspect_ratio="16:9",
    boundary=0.900,
    boundary_step_index=2,
    denoising_step_list=[1000, 750, 500, 250],
    config_json=None,
    rope_type="torch",
    resize_mode=None,
):
    self.resize_mode = resize_mode
    if config_json is not None:
        self.set_infer_config_json(config_json)
    else:
        self.set_infer_config(
            attn_mode,
            rope_type,
            infer_steps,
            num_frames,
            height,
            width,
            guidance_scale,
            sample_shift,
            fps,
            aspect_ratio,
            boundary,
            boundary_step_index,
            denoising_step_list,
        )

    config = set_config(self)
    print_config(config)
    self.runner = self._init_runner(config)  <-- Re-initialize the runner (model) again!!!

This is correct behavior from an implementation standpoint, but it is inefficient in practice since many of these parameters are generation-only and do not require model weight reloading.