| import torch, sys |
| sys.path.insert(0, r'C:\Users\Admin\Quillan-Ronin') |
| from transplant_weights import QuillanConfig, QuillanRoninSovereign |
|
|
| ckpt = torch.load('checkpoint_phase5.pt', weights_only=False, map_location='cpu') |
| cfg = ckpt['config'] |
| model = QuillanRoninSovereign(cfg) |
| model.load_state_dict(ckpt['model_state_dict'], strict=False) |
|
|
| dl = model.diffusion[0] |
| print(f'norm1.weight: [{dl.norm1.weight.min():.4f}, {dl.norm1.weight.max():.4f}]') |
| print(f'q_proj.weight: [{dl.q_proj.weight.min():.4f}, {dl.q_proj.weight.max():.4f}]') |
| print(f'k_proj.weight: [{dl.k_proj.weight.min():.4f}, {dl.k_proj.weight.max():.4f}]') |
| print(f'v_proj.weight: [{dl.v_proj.weight.min():.4f}, {dl.v_proj.weight.max():.4f}]') |
| print(f'o_proj.weight: [{dl.o_proj.weight.min():.4f}, {dl.o_proj.weight.max():.4f}]') |
|
|
| |
| for name, p in dl.ffn.named_parameters(): |
| print(f'ffn.{name}: [{p.min():.4f}, {p.max():.4f}]') |
|
|
| |
| print('\nAll diffusion layers:') |
| for i, layer in enumerate(model.diffusion): |
| print(f' Layer {i}: q_proj range=[{layer.q_proj.weight.min():.4f}, {layer.q_proj.weight.max():.4f}]') |
|
|
