Upload convert_custom_lora.py with huggingface_hub
Browse files- convert_custom_lora.py +117 -0
convert_custom_lora.py
ADDED
|
@@ -0,0 +1,117 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import argparse
|
| 2 |
+
import torch
|
| 3 |
+
from safetensors.torch import load_file, save_file
|
| 4 |
+
from collections import defaultdict
|
| 5 |
+
|
| 6 |
+
def convert_comfy_to_wan_lora_final_fp16(lora_path, output_path):
|
| 7 |
+
"""
|
| 8 |
+
Converts a ComfyUI-style LoRA to the format expected by 'wan.modules.model'.
|
| 9 |
+
- Keeps 'diffusion_model.' prefix.
|
| 10 |
+
- Converts 'lora_A' to 'lora_down', 'lora_B' to 'lora_up'.
|
| 11 |
+
- Skips per-layer '.alpha' keys.
|
| 12 |
+
- Skips keys related to 'img_emb.' that are under the 'diffusion_model.' prefix.
|
| 13 |
+
- Converts all LoRA weight tensors to float16.
|
| 14 |
+
|
| 15 |
+
Args:
|
| 16 |
+
lora_path (str): Path to the input ComfyUI LoRA .safetensors file.
|
| 17 |
+
output_path (str): Path to save the converted LoRA .safetensors file.
|
| 18 |
+
"""
|
| 19 |
+
try:
|
| 20 |
+
source_state_dict = load_file(lora_path)
|
| 21 |
+
except Exception as e:
|
| 22 |
+
print(f"Error loading LoRA file '{lora_path}': {e}")
|
| 23 |
+
return
|
| 24 |
+
|
| 25 |
+
diffusers_state_dict = {}
|
| 26 |
+
print(f"Loaded {len(source_state_dict)} tensors from {lora_path}")
|
| 27 |
+
|
| 28 |
+
source_comfy_prefix = "diffusion_model."
|
| 29 |
+
target_wan_prefix = "diffusion_model."
|
| 30 |
+
|
| 31 |
+
converted_count = 0
|
| 32 |
+
skipped_alpha_keys_count = 0
|
| 33 |
+
skipped_img_emb_keys_count = 0
|
| 34 |
+
problematic_keys = []
|
| 35 |
+
|
| 36 |
+
for key, tensor in source_state_dict.items():
|
| 37 |
+
original_key = key
|
| 38 |
+
|
| 39 |
+
if not key.startswith(source_comfy_prefix):
|
| 40 |
+
problematic_keys.append(f"{original_key} (Key does not start with expected prefix '{source_comfy_prefix}')")
|
| 41 |
+
continue
|
| 42 |
+
|
| 43 |
+
module_and_lora_part = key[len(source_comfy_prefix):]
|
| 44 |
+
|
| 45 |
+
if module_and_lora_part.startswith("img_emb."):
|
| 46 |
+
skipped_img_emb_keys_count += 1
|
| 47 |
+
continue
|
| 48 |
+
|
| 49 |
+
new_key_module_base = ""
|
| 50 |
+
new_lora_suffix = ""
|
| 51 |
+
is_weight_tensor = False # Flag to identify tensors that need dtype conversion
|
| 52 |
+
|
| 53 |
+
if module_and_lora_part.endswith(".lora_A.weight"):
|
| 54 |
+
new_key_module_base = module_and_lora_part[:-len(".lora_A.weight")]
|
| 55 |
+
new_lora_suffix = ".lora_down.weight"
|
| 56 |
+
is_weight_tensor = True
|
| 57 |
+
elif module_and_lora_part.endswith(".lora_B.weight"):
|
| 58 |
+
new_key_module_base = module_and_lora_part[:-len(".lora_B.weight")]
|
| 59 |
+
new_lora_suffix = ".lora_up.weight"
|
| 60 |
+
is_weight_tensor = True
|
| 61 |
+
elif module_and_lora_part.endswith(".alpha"):
|
| 62 |
+
skipped_alpha_keys_count += 1
|
| 63 |
+
continue # Alpha keys are skipped and don't need dtype conversion if they were kept
|
| 64 |
+
else:
|
| 65 |
+
problematic_keys.append(f"{original_key} (Unknown LoRA suffix or non-LoRA key within '{source_comfy_prefix}' structure: '...{module_and_lora_part[-25:]}')")
|
| 66 |
+
continue
|
| 67 |
+
|
| 68 |
+
new_key = target_wan_prefix + new_key_module_base + new_lora_suffix
|
| 69 |
+
|
| 70 |
+
# Convert to float16 if it's a weight tensor
|
| 71 |
+
if is_weight_tensor:
|
| 72 |
+
if tensor.is_floating_point(): # Only convert floating point types
|
| 73 |
+
diffusers_state_dict[new_key] = tensor.to(torch.float16)
|
| 74 |
+
else: # Should not happen for LoRA weights, but as a safeguard
|
| 75 |
+
diffusers_state_dict[new_key] = tensor
|
| 76 |
+
print(f"Warning: Tensor {original_key} was not floating point, dtype not changed.")
|
| 77 |
+
|
| 78 |
+
else: # Should not be reached if only lora_A/B weights are processed
|
| 79 |
+
diffusers_state_dict[new_key] = tensor
|
| 80 |
+
|
| 81 |
+
|
| 82 |
+
converted_count += 1
|
| 83 |
+
|
| 84 |
+
print(f"\nKey conversion finished.")
|
| 85 |
+
print(f"Successfully processed and converted {converted_count} LoRA weight keys (to float16).")
|
| 86 |
+
if skipped_alpha_keys_count > 0:
|
| 87 |
+
print(f"Skipped {skipped_alpha_keys_count} '.alpha' keys.")
|
| 88 |
+
if skipped_img_emb_keys_count > 0:
|
| 89 |
+
print(f"Skipped {skipped_img_emb_keys_count} 'diffusion_model.img_emb.' related keys.")
|
| 90 |
+
if problematic_keys:
|
| 91 |
+
print(f"Found {len(problematic_keys)} other keys that were also skipped (see details below):")
|
| 92 |
+
for pkey in problematic_keys:
|
| 93 |
+
print(f" - {pkey}")
|
| 94 |
+
|
| 95 |
+
if diffusers_state_dict:
|
| 96 |
+
print(f"Output dictionary has {len(diffusers_state_dict)} keys.")
|
| 97 |
+
print(f"Now attempting to save the file to: {output_path} (This might take a while for large files)...")
|
| 98 |
+
try:
|
| 99 |
+
save_file(diffusers_state_dict, output_path)
|
| 100 |
+
print(f"\nSuccessfully saved converted LoRA to: {output_path}")
|
| 101 |
+
except Exception as e:
|
| 102 |
+
print(f"Error saving converted LoRA file '{output_path}': {e}")
|
| 103 |
+
elif converted_count == 0 and source_state_dict:
|
| 104 |
+
print("\nNo keys were converted. Check input LoRA format and skipped key counts.")
|
| 105 |
+
elif not source_state_dict:
|
| 106 |
+
print("\nInput LoRA file seems empty or could not be loaded. No conversion performed.")
|
| 107 |
+
|
| 108 |
+
if __name__ == "__main__":
|
| 109 |
+
parser = argparse.ArgumentParser(
|
| 110 |
+
description="Convert ComfyUI-style LoRA to 'wan.modules.model' format, converting weights to float16.",
|
| 111 |
+
formatter_class=argparse.RawTextHelpFormatter
|
| 112 |
+
)
|
| 113 |
+
parser.add_argument("lora_path", type=str, help="Path to the input ComfyUI LoRA (.safetensors) file.")
|
| 114 |
+
parser.add_argument("output_path", type=str, help="Path to save the converted LoRA (.safetensors) file.")
|
| 115 |
+
args = parser.parse_args()
|
| 116 |
+
|
| 117 |
+
convert_comfy_to_wan_lora_final_fp16(args.lora_path, args.output_path)
|