| """Project configuration, not a Transformers AutoConfig or base-model config.""" |
| import json,math |
| from pathlib import Path |
| DEFAULT_CONFIG=Path(__file__).resolve().parents[2]/'config.json' |
| def load_config(path=None): |
| c=json.loads(Path(path or DEFAULT_CONFIG).read_text()) |
| if c['schema_version']!=1:raise ValueError('Unsupported config schema') |
| i,f=c['input'],c['frontend'] |
| if any(type(i[k]) is not int or i[k]<=0 for k in ['frames','height','width']):raise ValueError('Invalid input dimensions') |
| if i['height']*i['width']%8:raise ValueError('Frame dimensions must be byte aligned') |
| if i['packed_bytes']!=i['frames']*i['height']*i['width']//8:raise ValueError('packed_bytes mismatch') |
| if i['bitorder'] not in ['little','big'] or type(i['flip_height']) is not bool:raise ValueError('Invalid packing/orientation') |
| ch=f['channels'] |
| if len(ch)<2 or any(type(n) is not int or n<=0 for n in ch) or ch[0]!=i['frames'] or ch[-1]!=1:raise ValueError('Invalid frontend channels') |
| if f['architecture']!='Conv2dReLUStack' or f['activation']!='relu':raise ValueError('Unsupported frontend architecture') |
| if type(f['kernel_size']) is not int or f['kernel_size']<=0 or f['kernel_size']%2!=1 or f['stride']!=1 or f['padding']!=f['kernel_size']//2:raise ValueError('Only shape-preserving convolutions supported') |
| if f['output_clamp']!=[0.,1.]:raise ValueError('Output must use [0,1] scale') |
| g=c['generation']['inference'] |
| if not isinstance(g['prompt'],str) or not g['prompt'].strip():raise ValueError('Empty prompt') |
| if type(g['steps']) is not int or g['steps']<=0 or type(g['seed']) is not int:raise ValueError('Invalid steps/seed') |
| for k in ['cfg_text_scale','cfg_img_scale','timestep_shift','cfg_renorm_min']: |
| if not math.isfinite(g[k]):raise ValueError('Nonfinite inference parameter') |
| return c |
|
|
| def load_frontend_weights(path): |
| path=Path(path) |
| if path.suffix=='.safetensors': |
| from safetensors.torch import load_file |
| return load_file(str(path),device='cpu') |
| import torch |
| value=torch.load(path,map_location='cpu',weights_only=True) |
| return value.get('model',value) |
|
|
