File size: 3,020 Bytes
42cd39c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Reusable actual FP32 VAE encoding entry; Root runs only on its leased GPU."""
import argparse
import importlib.util
import json
from pathlib import Path
import sys
import numpy as np
from float_fields import load_video,normalize_for_vae


def load_official_vae(package,weight_file,device='cuda'):
    import torch
    torch.backends.cuda.matmul.allow_tf32=False
    torch.backends.cudnn.allow_tf32=False
    torch.backends.cudnn.benchmark=False
    torch.backends.cudnn.deterministic=True
    torch.set_float32_matmul_precision('highest')
    path=Path(package)/'code/LumiLens/diffsynth/models/wan_video_vae.py'
    spec=importlib.util.spec_from_file_location('official_vae38_float_entry',path)
    m=importlib.util.module_from_spec(spec);sys.modules[spec.name]=m;spec.loader.exec_module(m)
    vae=m.WanVideoVAE38()
    state=torch.load(weight_file,map_location='cpu',weights_only=True,mmap=True)
    vae.load_state_dict(vae.state_dict_converter().from_civitai(state),strict=True)
    del state
    vae=vae.to(device=device,dtype=torch.float32).eval().requires_grad_(False)
    assert all(x.dtype==torch.float32 for x in vae.parameters())
    return vae


def encode_array_fp32(vae,array,device='cuda'):
    import torch
    assert all(x.dtype==torch.float32 for x in vae.parameters()),'Reject premature reduced-precision VAE'
    assert array.dtype==np.float32 and array.ndim==4 and array.shape[-1]==3
    assert (len(array)-1)%4==0
    torch.backends.cuda.matmul.allow_tf32=False
    torch.backends.cudnn.allow_tf32=False
    v=torch.from_numpy(normalize_for_vae(array)).permute(3,0,1,2)[None].to(device,dtype=torch.float32)
    with torch.inference_mode(),torch.autocast(device_type='cuda',enabled=False):
        z=vae.encode(v,device=device,tiled=False).cpu()
    assert z.dtype==torch.float32 and torch.isfinite(z).all()
    assert z.shape[1:]==(48,(len(array)-1)//4+1,array.shape[1]//16,array.shape[2]//16)
    return z


def main():
    p=argparse.ArgumentParser();p.add_argument('--task',required=True);p.add_argument('--vae-file',required=True);p.add_argument('--out',required=True);p.add_argument('--package',default='/mnt/atlas/fangrui/LUMILENS_QINGHAI_TRAINING_20261010');a=p.parse_args()
    import torch
    assert torch.cuda.is_available(),'Root-owned GPU stage required'
    out=Path(a.out);assert not out.exists();out.mkdir(parents=True)
    task=json.loads(Path(a.task).read_text());vae=load_official_vae(a.package,a.vae_file)
    for field in ('color','intensity'):
        array=load_video(task['float_fields'],field)
        z=encode_array_fp32(vae,array)
        torch.save(z,out/(field+'_fp32_full49.pt'))
    (out/'receipt.json').write_text(json.dumps({'schema':'lampedit.fp32_condition_encoding_entry.v001','sample_id':task['sample_id'],'frame_count':49,'latent_dtype':'float32','no_uint8_or_BF16_light_path':True,'ready_for_training':False,'scope':'two physical light conditions only; source/target/joint/text/adapter acceptance separate'},indent=2)+'\n')


if __name__=='__main__':main()