SwiftFusion/inference.py at release · OpenImagingLab/SwiftFusion · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
import argparse
from pathlib import Path

import torch
from tqdm import tqdm

from models.swiftfusion_checkpoints import (
    model_configs_from_directory,
    tokenizer_config_from_directory,
)
from pipelines.wan_video_new import WanVideoPipeline
from prompts import DEFAULT_PROMPT
from trainers.unified_dataset import UnifiedDataset


def parse_args():
    parser = argparse.ArgumentParser(
        description="Run block-sparse SwiftFusion v1.38 inference."
    )
    parser.add_argument("--metadata_path", required=True)
    parser.add_argument("--model_dir", required=True)
    parser.add_argument("--raft_checkpoint", required=True)
    parser.add_argument("--base_lora_checkpoint", required=True)
    parser.add_argument("--distilled_lora_checkpoint", required=True)
    parser.add_argument("--sparse_lora_checkpoint", required=True)
    parser.add_argument("--tokenizer_dir", default=None)
    parser.add_argument("--data_root", default="")
    parser.add_argument("--output_dir", default="./outputs/inference")
    parser.add_argument("--max_pixels", type=int, default=512 * 640)
    parser.add_argument("--num_inference_steps", type=int, default=1)
    parser.add_argument("--seed", type=int, default=1)
    parser.add_argument("--num_workers", type=int, default=8)
    return parser.parse_args()


def require_file(path, label):
    path = Path(path)
    if not path.is_file():
        raise FileNotFoundError(f"{label} not found: {path}")


def main():
    args = parse_args()
    checkpoints = (
        (args.metadata_path, "Metadata"),
        (args.raft_checkpoint, "RAFT checkpoint"),
        (args.base_lora_checkpoint, "Stage 1 checkpoint"),
        (
            args.distilled_lora_checkpoint,
            "Stage 3 checkpoint",
        ),
        (args.sparse_lora_checkpoint, "Stage 4 checkpoint"),
    )
    for path, label in checkpoints:
        require_file(path, label)

    pipe = WanVideoPipeline.from_pretrained(
        torch_dtype=torch.bfloat16,
        device="cuda",
        model_configs=model_configs_from_directory(
            args.model_dir,
            offload_device="cpu",
        ),
        tokenizer_config=tokenizer_config_from_directory(
            args.model_dir,
            args.tokenizer_dir,
        ),
        raft_checkpoint=args.raft_checkpoint,
    )
    pipe.dit.enable_block_sparse_attention(
        layer_stride=2,
    )
    pipe.load_lora(
        pipe.dit,
        args.base_lora_checkpoint,
        alpha=1.0,
    )
    pipe.load_lora(
        pipe.dit,
        args.distilled_lora_checkpoint,
        alpha=1.0,
        remove_prefix="pipe.dit.",
        delete_prefix="fake_teacher.dit.",
    )
    pipe.load_lora(
        pipe.dit,
        args.sparse_lora_checkpoint,
        alpha=1.0,
    )
    pipe.eval()
    pipe.enable_vram_management()

    dataset = UnifiedDataset(
        base_path=args.data_root,
        metadata_path=args.metadata_path,
        repeat=1,
        data_file_keys=("gt", "video"),
        main_data_operator=UnifiedDataset.default_image_operator(
            base_path=args.data_root,
            max_pixels=args.max_pixels,
            height_division_factor=16,
            width_division_factor=16,
        ),
    )
    dataloader = torch.utils.data.DataLoader(
        dataset,
        shuffle=False,
        collate_fn=lambda batch: batch[0],
        num_workers=args.num_workers,
    )
    output_dir = Path(args.output_dir)
    output_dir.mkdir(parents=True, exist_ok=True)

    with torch.inference_mode():
        for step, data in enumerate(tqdm(dataloader)):
            input_video = data["video"]
            output, _, _ = pipe(
                prompt=data.get("prompt", DEFAULT_PROMPT),
                negative_prompt="",
                input_video=input_video,
                height=input_video[0].height,
                width=input_video[0].width,
                num_frames=2,
                cfg_scale=1,
                seed=args.seed,
                num_inference_steps=args.num_inference_steps,
            progress_bar_cmd=lambda timesteps: timesteps,
            )
            result = pipe.vae_output_to_video(output)[0]
            input_video[0].save(output_dir / f"{step}_oe.png")
            input_video[1].save(output_dir / f"{step}_ue.png")
            if "gt" in data:
                data["gt"].save(output_dir / f"{step}_gt.png")
            result.save(output_dir / f"{step}_result.png")


if __name__ == "__main__":
    main()