File size: 6,001 Bytes
7d72aa0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
"""
Prepare inference code + checkpoints for the Soundstage Space.

Modes:
1) Monorepo: if parent folder already has inference_v2st.py (this project), use it.
2) Standalone Space: clone Foley-Omni from GitHub, then download Hub weights.
"""

from __future__ import annotations

import os
import subprocess
from pathlib import Path

HERE = Path(__file__).resolve().parent
PARENT = HERE.parent

REPO_URL = os.getenv("FOLEY_REPO_URL", "https://github.com/SAIL-PRLab/Foley-Omni.git").strip()
REPO_COMMIT = os.getenv("FOLEY_COMMIT", "").strip()
WEIGHTS_REPO = os.getenv("FOLEY_WEIGHTS_REPO", "CocoBro/Foley-Omni").strip()

# When Space repo is only hf_space/*, code lives under vendor/
VENDOR_DIR = HERE / "vendor" / "Foley-Omni"

CKPT_FILES = [
    "ckpts/Foley-Omni/v2st.pth",
    "ckpts/Wan2.2-TI2V-5B/models_t5_umt5-xxl-enc-bf16.pth",
    "ckpts/Wan2.2-TI2V-5B/google/umt5-xxl/special_tokens_map.json",
    "ckpts/Wan2.2-TI2V-5B/google/umt5-xxl/spiece.model",
    "ckpts/Wan2.2-TI2V-5B/google/umt5-xxl/tokenizer.json",
    "ckpts/Wan2.2-TI2V-5B/google/umt5-xxl/tokenizer_config.json",
    "ckpts/mmaudio/ext_weights/v1-16.pth",
    "ckpts/mmaudio/ext_weights/best_netG.pt",
    "ckpts/mmaudio/ext_weights/synchformer_state_dict.pth",
]

REPO_DIR: Path | None = None


def resolve_repo_dir() -> Path:
    global REPO_DIR
    if REPO_DIR is not None:
        return REPO_DIR

    if (PARENT / "inference_v2st.py").is_file() and (PARENT / "foley_omni").is_dir():
        REPO_DIR = PARENT
        print(f"[setup] using monorepo at {REPO_DIR}")
        return REPO_DIR

    REPO_DIR = VENDOR_DIR
    print(f"[setup] using vendor dir {REPO_DIR}")
    return REPO_DIR


def _apply_patch(file_path: Path, old: str, new: str, label: str) -> None:
    content = file_path.read_text(encoding="utf-8")
    if new in content:
        print(f"[patch] {label}: already patched")
        return
    if old not in content:
        print(f"[patch] {label}: pattern not found — skipping (upstream may differ)")
        return
    file_path.write_text(content.replace(old, new), encoding="utf-8")
    print(f"[patch] {label}: patched")


def apply_zerogpu_patches(repo: Path) -> None:
    """Best-effort RAM / attention patches used by the working FOG Space."""
    m = repo / "foley_omni"
    if not m.is_dir():
        return

    _apply_patch(
        m / "modules" / "t5.py",
        old="""    # init model
        model = umt5_xxl(
            encoder_only=True,
            return_tokenizer=False,
            dtype=dtype,
            device=device if not cpu_offload else "cpu").eval().requires_grad_(False)
        logging.info(f'loading {checkpoint_path}')
        model.load_state_dict(torch.load(checkpoint_path, map_location='cpu'))
        self.model = model""",
        new="""    # init model on the meta device (allocates no real memory yet)
        model = umt5_xxl(
            encoder_only=True,
            return_tokenizer=False,
            dtype=dtype,
            device="meta").eval().requires_grad_(False)
        logging.info(f'loading {checkpoint_path}')
        try:
            state_dict = torch.load(checkpoint_path, map_location='cpu', mmap=True, weights_only=True)
        except Exception:
            state_dict = torch.load(checkpoint_path, map_location='cpu')
        model.load_state_dict(state_dict, assign=True)
        del state_dict
        target_device = device if not cpu_offload else "cpu"
        model = model.to(target_device)
        self.model = model""",
        label="t5.py",
    )

    _apply_patch(
        m / "fusion_engine.py",
        old="""            logging.info(f"Loading model checkpoint from {checkpoint_path}")
            checkpoint = torch.load(checkpoint_path, map_location="cpu")
            if isinstance(checkpoint, dict) and "model" in checkpoint:""",
        new="""            logging.info(f"Loading model checkpoint from {checkpoint_path}")
            try:
                checkpoint = torch.load(checkpoint_path, map_location="cpu", mmap=True, weights_only=True)
            except Exception:
                checkpoint = torch.load(checkpoint_path, map_location="cpu")
            if isinstance(checkpoint, dict) and "model" in checkpoint:""",
        label="fusion_engine.py",
    )

    imp_old = "from .attention import flash_attention"
    imp_new = "from .attention import attention as flash_attention"
    for rel in ("modules/model.py", "modules/clip.py", "modules/__init__.py"):
        p = m / rel
        if p.is_file():
            _apply_patch(p, imp_old, imp_new, f"{rel} attention import")


def clone_repo_if_needed(repo: Path) -> None:
    if (repo / "inference_v2st.py").is_file():
        print("[setup] inference code already present")
        return
    repo.parent.mkdir(parents=True, exist_ok=True)
    if REPO_COMMIT:
        subprocess.run(["git", "clone", REPO_URL, str(repo)], check=True)
        subprocess.run(["git", "-C", str(repo), "checkout", REPO_COMMIT], check=True)
    else:
        subprocess.run(["git", "clone", "--depth", "1", REPO_URL, str(repo)], check=True)


def download_checkpoints(repo: Path) -> None:
    missing = [f for f in CKPT_FILES if not (repo / f).is_file()]
    if not missing:
        print("[setup] checkpoints already present")
        return
    from huggingface_hub import snapshot_download

    print(f"[setup] downloading {len(missing)} files from {WEIGHTS_REPO} ...")
    token = os.getenv("HF_TOKEN") or os.getenv("HUGGING_FACE_HUB_TOKEN")
    snapshot_download(
        WEIGHTS_REPO,
        local_dir=str(repo),
        allow_patterns=CKPT_FILES,
        token=token,
    )
    still = [f for f in CKPT_FILES if not (repo / f).is_file()]
    if still:
        raise RuntimeError(f"Checkpoint download incomplete: {still}")


def ensure_ready() -> Path:
    repo = resolve_repo_dir()
    clone_repo_if_needed(repo)
    apply_zerogpu_patches(repo)
    download_checkpoints(repo)
    print("[setup] ready")
    return repo


if __name__ == "__main__":
    ensure_ready()