create_lora.py

"""
Adapted from 
https://github.com/Stability-AI/stability-ComfyUI-nodes/blob/001154622564b17223ce0191803c5fff7b87146c/control_lora_create.py
"""

from diffusers import CogVideoXTransformer3DModel
from tqdm.auto import tqdm
from safetensors.torch import save_file
import torch

RANK = 64
CLAMP_QUANTILE = 0.99


# Comes from
# https://github.com/Stability-AI/stability-ComfyUI-nodes/blob/001154622564b17223ce0191803c5fff7b87146c/control_lora_create.py#L9
def extract_lora(diff, rank):
    if torch.cuda.is_available():
        diff = diff.to("cuda")
    
    is_conv2d = (len(diff.shape) == 4)
    kernel_size = None if not is_conv2d else diff.size()[2:4]
    is_conv2d_3x3 = is_conv2d and kernel_size != (1, 1)
    out_dim, in_dim = diff.size()[0:2]
    rank = min(rank, in_dim, out_dim)

    if is_conv2d:
        if is_conv2d_3x3:
            diff = diff.flatten(start_dim=1)
        else:
            diff = diff.squeeze()

    U, S, Vh = torch.linalg.svd(diff.float())
    U = U[:, :rank]
    S = S[:rank]
    U = U @ torch.diag(S)
    Vh = Vh[:rank, :]

    dist = torch.cat([U.flatten(), Vh.flatten()])
    hi_val = torch.quantile(dist, CLAMP_QUANTILE)
    low_val = -hi_val

    U = U.clamp(low_val, hi_val)
    Vh = Vh.clamp(low_val, hi_val)
    if is_conv2d:
        U = U.reshape(out_dim, rank, 1, 1)
        Vh = Vh.reshape(rank, in_dim, kernel_size[0], kernel_size[1])
    return (U.cpu(), Vh.cpu())


transformer_finetuned = CogVideoXTransformer3DModel.from_pretrained(
    "cogvideox-cakeify", subfolder="transformer", torch_dtype=torch.bfloat16
)
state_dict_ft = transformer_finetuned.state_dict()

transformer = CogVideoXTransformer3DModel.from_pretrained(
    "THUDM/CogVideoX-5b", subfolder="transformer", torch_dtype=torch.bfloat16
)
state_dict = transformer.state_dict()
output_dict = {}

for k in tqdm(state_dict, desc="Extracting LoRA..."):
    original_param = state_dict[k]
    finetuned_param = state_dict_ft[k]
    if len(original_param.shape) >= 2:
        diff = finetuned_param.float() - original_param.float()
        out = extract_lora(diff, RANK)
        name = k
        
        if name.endswith(".weight"):
            name = name[:-len(".weight")]
        down_key = "{}.lora_A.weight".format(name)
        up_key = "{}.lora_B.weight".format(name)
        
        output_dict[up_key] = out[0].contiguous().to(finetuned_param.dtype)
        output_dict[down_key] = out[1].contiguous().to(finetuned_param.dtype)

output_dict = {f"transformer.{k}": v for k, v in output_dict.items()}
save_file(output_dict, "extracted_cakeify_lora_64.safetensors")
print(f"LoRA saved and it contains {len(output_dict)} keys.")