methods/sparse_attention/Sparse-VideoGen/cog_inference.py

import argparse

import torch
from diffusers import CogVideoXImageToVideoPipeline

from svg.models.cog.inference import replace_cog_attention, sample_image
from svg.models.cog.utils import seed_everything

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="A script that sets a random seed.")
    parser.add_argument("--version", type=str, default="v1.5", choices=["v1", "v1.5"], help="Random seed for reproducibility")
    parser.add_argument("--seed", type=int, default=0, help="Random seed for reproducibility")
    parser.add_argument("--image_path", type=str, required=True, help="Image Path")
    parser.add_argument("--prompt", type=str, required=True, help="Prompt")
    parser.add_argument("--pattern", type=str, default="SVG", choices=["SVG", "dense"])
    parser.add_argument("--num_step", type=int, default=50, help="Number of steps to inference")
    parser.add_argument("--first_layers_fp", type=float, default=0.025, help="Only works for best config. Leave the 0, 1, 2, 40, 41 layers in FP")
    parser.add_argument("--first_times_fp", type=float, default=0.2, help="Only works for best config. Leave the first 10% timestep in FP")
    parser.add_argument("--num_sampled_rows", type=int, default=32, help="The number of sampled rows")
    parser.add_argument("--sparsity", type=float, default=0.25, help="The sparsity of the striped attention pattern. Accepts one or two float values. Only effective for fast_sample_mse")
    parser.add_argument("--output_path", type=str, required=True, help="Output generated videos")

    args = parser.parse_args()

    seed_everything(args.seed)

    model_id = "THUDM/CogVideoX1.5-5B-I2V"

    dtype = torch.bfloat16

    pipe = CogVideoXImageToVideoPipeline.from_pretrained(model_id, torch_dtype=dtype).to("cuda")

    pipe.vae.enable_tiling()
    pipe.vae.enable_slicing()

    if args.pattern == "SVG":
        replace_cog_attention(pipe, args.version, args.num_sampled_rows, args.sparsity, args.first_layers_fp, args.first_times_fp)

    sample_image(pipe, args.prompt, args.image_path, args.output_path, args.seed, args.version, args.num_step)