SD1.5 Unet from_single_file loading does not work

[AbhinavGopal]

Describe the bug

SD1.5 Unet from_single_file loading does not work (either from safetensor or GGUF)

Reproduction

`
import torch
from diffusers import UNet2DConditionModel

config = UNet2DConditionModel.load_config("SimianLuo/LCM_Dreamshaper_v7", subfolder="unet")
unet = UNet2DConditionModel.from_single_file(
"stable-diffusion.cpp/build/diffusion_pytorch_model.safetensors",
config="SimianLuo/LCM_Dreamshaper_v7", # Use the repo ID string
subfolder='unet',
torch_dtype=torch.bfloat16,
use_safetensors=True
)
`

and
`
import torch
from diffusers import UNet2DConditionModel

config = UNet2DConditionModel.load_config("SimianLuo/LCM_Dreamshaper_v7", subfolder="unet")
unet = UNet2DConditionModel.from_single_file(
"https://huggingface.co/abhinavgopal/firstfile.gguf/LCM_Dreamshaper_v7_4k.safetensors.q8_0.gguf",
config="SimianLuo/LCM_Dreamshaper_v7", # Use the repo ID string
subfolder='unet',
torch_dtype=torch.bfloat16,
use_safetensors=True
)
`

Logs

Traceback (most recent call last):
  File "/home/ec2-user/temp.py", line 14, in <module>
    unet = UNet2DConditionModel.from_single_file(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ec2-user/miniconda3/lib/python3.12/site-packages/huggingface_hub/utils/_validators.py", line 114, in _inner_fn
    return fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^
  File "/home/ec2-user/miniconda3/lib/python3.12/site-packages/diffusers/loaders/single_file_model.py", line 325, in from_single_file
    raise SingleFileComponentError(
diffusers.loaders.single_file_utils.SingleFileComponentError: Failed to load UNet2DConditionModel. Weights for this component appear to be missing in the checkpoint.

System Info

• 🤗 Diffusers version: 0.32.2
• Platform: Linux-6.1.128-136.201.amzn2023.x86_64-x86_64-with-glibc2.34
• Running on Google Colab?: No
• Python version: 3.12.9
• PyTorch version (GPU?): 2.6.0+cu124 (True)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Huggingface_hub version: 0.29.2
• Transformers version: 4.49.0
• Accelerate version: 1.4.0
• PEFT version: not installed
• Bitsandbytes version: 0.45.3
• Safetensors version: 0.5.3
• xFormers version: not installed
• Accelerator: NVIDIA L40S, 46068 MiB
• Using GPU in script?: No
• Using distributed or parallel set-up in script?: No

Who can help?

@yiyixuxu @asomoza @sayakpaul

[AbhinavGopal]

additional info: the unet here https://github.com/huggingface/diffusers/blob/main/src/diffusers/loaders/single_file_utils.py#L1062 never gets populated, and the keys https://github.com/huggingface/diffusers/blob/main/src/diffusers/loaders/single_file_utils.py#L1063 are the following:

['conv_in.bias', 'conv_in.weight', 'conv_norm_out.bias', 'conv_norm_out.weight', 'conv_out.bias', 'conv_out.weight', 'down_blocks.0.attentions.0.norm.bias', 'down_blocks.0.attentions.0.norm.weight', 'down_blocks.0.attentions.0.proj_in.bias', 'down_blocks.0.attentions.0.proj_in.weight', 'down_blocks.0.attentions.0.proj_out.bias', 'down_blocks.0.attentions.0.proj_out.weight', 'down_blocks.0.attentions.0.transformer_blocks.0.attn1.to_k.weight', 'down_blocks.0.attentions.0.transformer_blocks.0.attn1.to_out.0.bias', 'down_blocks.0.attentions.0.transformer_blocks.0.attn1.to_out.0.weight', 'down_blocks.0.attentions.0.transformer_blocks.0.attn1.to_q.weight', 'down_blocks.0.attentions.0.transformer_blocks.0.attn1.to_v.weight', 'down_blocks.0.attentions.0.transformer_blocks.0.attn2.to_k.weight', 'down_blocks.0.attentions.0.transformer_blocks.0.attn2.to_out.0.bias', 'down_blocks.0.attentions.0.transformer_blocks.0.attn2.to_out.0.weight', 'down_blocks.0.attentions.0.transformer_blocks.0.attn2.to_q.weight', 'down_blocks.0.attentions.0.transformer_blocks.0.attn2.to_v.weight', 'down_blocks.0.attentions.0.transformer_blocks.0.ff.net.0.proj.bias', 'down_blocks.0.attentions.0.transformer_blocks.0.ff.net.0.proj.weight', 'down_blocks.0.attentions.0.transformer_blocks.0.ff.net.2.bias', 'down_blocks.0.attentions.0.transformer_blocks.0.ff.net.2.weight', 'down_blocks.0.attentions.0.transformer_blocks.0.norm1.bias', 'down_blocks.0.attentions.0.transformer_blocks.0.norm1.weight', 'down_blocks.0.attentions.0.transformer_blocks.0.norm2.bias', 'down_blocks.0.attentions.0.transformer_blocks.0.norm2.weight', 'down_blocks.0.attentions.0.transformer_blocks.0.norm3.bias', 'down_blocks.0.attentions.0.transformer_blocks.0.norm3.weight', 'down_blocks.0.attentions.1.norm.bias', 'down_blocks.0.attentions.1.norm.weight', 'down_blocks.0.attentions.1.proj_in.bias', 'down_blocks.0.attentions.1.proj_in.weight', 'down_blocks.0.attentions.1.proj_out.bias', 'down_blocks.0.attentions.1.proj_out.weight', 'down_blocks.0.attentions.1.transformer_blocks.0.attn1.to_k.weight', 'down_blocks.0.attentions.1.transformer_blocks.0.attn1.to_out.0.bias', 'down_blocks.0.attentions.1.transformer_blocks.0.attn1.to_out.0.weight', 'down_blocks.0.attentions.1.transformer_blocks.0.attn1.to_q.weight', 'down_blocks.0.attentions.1.transformer_blocks.0.attn1.to_v.weight', 'down_blocks.0.attentions.1.transformer_blocks.0.attn2.to_k.weight', 'down_blocks.0.attentions.1.transformer_blocks.0.attn2.to_out.0.bias', 'down_blocks.0.attentions.1.transformer_blocks.0.attn2.to_out.0.weight', 'down_blocks.0.attentions.1.transformer_blocks.0.attn2.to_q.weight', 'down_blocks.0.attentions.1.transformer_blocks.0.attn2.to_v.weight', 'down_blocks.0.attentions.1.transformer_blocks.0.ff.net.0.proj.bias', 'down_blocks.0.attentions.1.transformer_blocks.0.ff.net.0.proj.weight', 'down_blocks.0.attentions.1.transformer_blocks.0.ff.net.2.bias', 'down_blocks.0.attentions.1.transformer_blocks.0.ff.net.2.weight', 'down_blocks.0.attentions.1.transformer_blocks.0.norm1.bias', 'down_blocks.0.attentions.1.transformer_blocks.0.norm1.weight', 'down_blocks.0.attentions.1.transformer_blocks.0.norm2.bias', 'down_blocks.0.attentions.1.transformer_blocks.0.norm2.weight', 'down_blocks.0.attentions.1.transformer_blocks.0.norm3.bias', 'down_blocks.0.attentions.1.transformer_blocks.0.norm3.weight', 'down_blocks.0.downsamplers.0.conv.bias', 'down_blocks.0.downsamplers.0.conv.weight', 'down_blocks.0.resnets.0.conv1.bias', 'down_blocks.0.resnets.0.conv1.weight', 'down_blocks.0.resnets.0.conv2.bias', 'down_blocks.0.resnets.0.conv2.weight', 'down_blocks.0.resnets.0.norm1.bias', 'down_blocks.0.resnets.0.norm1.weight', 'down_blocks.0.resnets.0.norm2.bias', 'down_blocks.0.resnets.0.norm2.weight', 'down_blocks.0.resnets.0.time_emb_proj.bias', 'down_blocks.0.resnets.0.time_emb_proj.weight', 'down_blocks.0.resnets.1.conv1.bias', 'down_blocks.0.resnets.1.conv1.weight', 'down_blocks.0.resnets.1.conv2.bias', 'down_blocks.0.resnets.1.conv2.weight', 'down_blocks.0.resnets.1.norm1.bias', 'down_blocks.0.resnets.1.norm1.weight', 'down_blocks.0.resnets.1.norm2.bias', 'down_blocks.0.resnets.1.norm2.weight', 'down_blocks.0.resnets.1.time_emb_proj.bias', 'down_blocks.0.resnets.1.time_emb_proj.weight', 'down_blocks.1.attentions.0.norm.bias', 'down_blocks.1.attentions.0.norm.weight', 'down_blocks.1.attentions.0.proj_in.bias', 'down_blocks.1.attentions.0.proj_in.weight', 'down_blocks.1.attentions.0.proj_out.bias', 'down_blocks.1.attentions.0.proj_out.weight', 'down_blocks.1.attentions.0.transformer_blocks.0.attn1.to_k.weight', 'down_blocks.1.attentions.0.transformer_blocks.0.attn1.to_out.0.bias', 'down_blocks.1.attentions.0.transformer_blocks.0.attn1.to_out.0.weight', 'down_blocks.1.attentions.0.transformer_blocks.0.attn1.to_q.weight', 'down_blocks.1.attentions.0.transformer_blocks.0.attn1.to_v.weight', 'down_blocks.1.attentions.0.transformer_blocks.0.attn2.to_k.weight', 'down_blocks.1.attentions.0.transformer_blocks.0.attn2.to_out.0.bias', 'down_blocks.1.attentions.0.transformer_blocks.0.attn2.to_out.0.weight', 'down_blocks.1.attentions.0.transformer_blocks.0.attn2.to_q.weight', 'down_blocks.1.attentions.0.transformer_blocks.0.attn2.to_v.weight', 'down_blocks.1.attentions.0.transformer_blocks.0.ff.net.0.proj.bias', 'down_blocks.1.attentions.0.transformer_blocks.0.ff.net.0.proj.weight', 'down_blocks.1.attentions.0.transformer_blocks.0.ff.net.2.bias', 'down_blocks.1.attentions.0.transformer_blocks.0.ff.net.2.weight', 'down_blocks.1.attentions.0.transformer_blocks.0.norm1.bias', 'down_blocks.1.attentions.0.transformer_blocks.0.norm1.weight', 'down_blocks.1.attentions.0.transformer_blocks.0.norm2.bias', 'down_blocks.1.attentions.0.transformer_blocks.0.norm2.weight', 'down_blocks.1.attentions.0.transformer_blocks.0.norm3.bias', 'down_blocks.1.attentions.0.transformer_blocks.0.norm3.weight', 'down_blocks.1.attentions.1.norm.bias', 'down_blocks.1.attentions.1.norm.weight', 'down_blocks.1.attentions.1.proj_in.bias', 'down_blocks.1.attentions.1.proj_in.weight', 'down_blocks.1.attentions.1.proj_out.bias', 'down_blocks.1.attentions.1.proj_out.weight', 'down_blocks.1.attentions.1.transformer_blocks.0.attn1.to_k.weight', 'down_blocks.1.attentions.1.transformer_blocks.0.attn1.to_out.0.bias', 'down_blocks.1.attentions.1.transformer_blocks.0.attn1.to_out.0.weight', 'down_blocks.1.attentions.1.transformer_blocks.0.attn1.to_q.weight', 'down_blocks.1.attentions.1.transformer_blocks.0.attn1.to_v.weight', 'down_blocks.1.attentions.1.transformer_blocks.0.attn2.to_k.weight', 'down_blocks.1.attentions.1.transformer_blocks.0.attn2.to_out.0.bias', 'down_blocks.1.attentions.1.transformer_blocks.0.attn2.to_out.0.weight', 'down_blocks.1.attentions.1.transformer_blocks.0.attn2.to_q.weight', 'down_blocks.1.attentions.1.transformer_blocks.0.attn2.to_v.weight', 'down_blocks.1.attentions.1.transformer_blocks.0.ff.net.0.proj.bias', 'down_blocks.1.attentions.1.transformer_blocks.0.ff.net.0.proj.weight', 'down_blocks.1.attentions.1.transformer_blocks.0.ff.net.2.bias', 'down_blocks.1.attentions.1.transformer_blocks.0.ff.net.2.weight', 'down_blocks.1.attentions.1.transformer_blocks.0.norm1.bias', 'down_blocks.1.attentions.1.transformer_blocks.0.norm1.weight', 'down_blocks.1.attentions.1.transformer_blocks.0.norm2.bias', 'down_blocks.1.attentions.1.transformer_blocks.0.norm2.weight', 'down_blocks.1.attentions.1.transformer_blocks.0.norm3.bias', 'down_blocks.1.attentions.1.transformer_blocks.0.norm3.weight', 'down_blocks.1.downsamplers.0.conv.bias', 'down_blocks.1.downsamplers.0.conv.weight', 'down_blocks.1.resnets.0.conv1.bias', 'down_blocks.1.resnets.0.conv1.weight', 'down_blocks.1.resnets.0.conv2.bias', 'down_blocks.1.resnets.0.conv2.weight', 'down_blocks.1.resnets.0.conv_shortcut.bias', 'down_blocks.1.resnets.0.conv_shortcut.weight', 'down_blocks.1.resnets.0.norm1.bias', 'down_blocks.1.resnets.0.norm1.weight', 'down_blocks.1.resnets.0.norm2.bias', 'down_blocks.1.resnets.0.norm2.weight', 'down_blocks.1.resnets.0.time_emb_proj.bias', 'down_blocks.1.resnets.0.time_emb_proj.weight', 'down_blocks.1.resnets.1.conv1.bias', 'down_blocks.1.resnets.1.conv1.weight', 'down_blocks.1.resnets.1.conv2.bias', 'down_blocks.1.resnets.1.conv2.weight', 'down_blocks.1.resnets.1.norm1.bias', 'down_blocks.1.resnets.1.norm1.weight', 'down_blocks.1.resnets.1.norm2.bias', 'down_blocks.1.resnets.1.norm2.weight', 'down_blocks.1.resnets.1.time_emb_proj.bias', 'down_blocks.1.resnets.1.time_emb_proj.weight', 'down_blocks.2.attentions.0.norm.bias', 'down_blocks.2.attentions.0.norm.weight', 'down_blocks.2.attentions.0.proj_in.bias', 'down_blocks.2.attentions.0.proj_in.weight', 'down_blocks.2.attentions.0.proj_out.bias', 'down_blocks.2.attentions.0.proj_out.weight', 'down_blocks.2.attentions.0.transformer_blocks.0.attn1.to_k.weight', 'down_blocks.2.attentions.0.transformer_blocks.0.attn1.to_out.0.bias', 'down_blocks.2.attentions.0.transformer_blocks.0.attn1.to_out.0.weight', 'down_blocks.2.attentions.0.transformer_blocks.0.attn1.to_q.weight', 'down_blocks.2.attentions.0.transformer_blocks.0.attn1.to_v.weight', 'down_blocks.2.attentions.0.transformer_blocks.0.attn2.to_k.weight', 'down_blocks.2.attentions.0.transformer_blocks.0.attn2.to_out.0.bias', 'down_blocks.2.attentions.0.transformer_blocks.0.attn2.to_out.0.weight', 'down_blocks.2.attentions.0.transformer_blocks.0.attn2.to_q.weight', 'down_blocks.2.attentions.0.transformer_blocks.0.attn2.to_v.weight', 'down_blocks.2.attentions.0.transformer_blocks.0.ff.net.0.proj.bias', 'down_blocks.2.attentions.0.transformer_blocks.0.ff.net.0.proj.weight', 'down_blocks.2.attentions.0.transformer_blocks.0.ff.net.2.bias', 'down_blocks.2.attentions.0.transformer_blocks.0.ff.net.2.weight', 'down_blocks.2.attentions.0.transformer_blocks.0.norm1.bias', 'down_blocks.2.attentions.0.transformer_blocks.0.norm1.weight', 'down_blocks.2.attentions.0.transformer_blocks.0.norm2.bias', 'down_blocks.2.attentions.0.transformer_blocks.0.norm2.weight', 'down_blocks.2.attentions.0.transformer_blocks.0.norm3.bias', 'down_blocks.2.attentions.0.transformer_blocks.0.norm3.weight', 'down_blocks.2.attentions.1.norm.bias', 'down_blocks.2.attentions.1.norm.weight', 'down_blocks.2.attentions.1.proj_in.bias', 'down_blocks.2.attentions.1.proj_in.weight', 'down_blocks.2.attentions.1.proj_out.bias', 'down_blocks.2.attentions.1.proj_out.weight', 'down_blocks.2.attentions.1.transformer_blocks.0.attn1.to_k.weight', 'down_blocks.2.attentions.1.transformer_blocks.0.attn1.to_out.0.bias', 'down_blocks.2.attentions.1.transformer_blocks.0.attn1.to_out.0.weight', 'down_blocks.2.attentions.1.transformer_blocks.0.attn1.to_q.weight', 'down_blocks.2.attentions.1.transformer_blocks.0.attn1.to_v.weight', 'down_blocks.2.attentions.1.transformer_blocks.0.attn2.to_k.weight', 'down_blocks.2.attentions.1.transformer_blocks.0.attn2.to_out.0.bias', 'down_blocks.2.attentions.1.transformer_blocks.0.attn2.to_out.0.weight', 'down_blocks.2.attentions.1.transformer_blocks.0.attn2.to_q.weight', 'down_blocks.2.attentions.1.transformer_blocks.0.attn2.to_v.weight', 'down_blocks.2.attentions.1.transformer_blocks.0.ff.net.0.proj.bias', 'down_blocks.2.attentions.1.transformer_blocks.0.ff.net.0.proj.weight', 'down_blocks.2.attentions.1.transformer_blocks.0.ff.net.2.bias', 'down_blocks.2.attentions.1.transformer_blocks.0.ff.net.2.weight', 'down_blocks.2.attentions.1.transformer_blocks.0.norm1.bias', 'down_blocks.2.attentions.1.transformer_blocks.0.norm1.weight', 'down_blocks.2.attentions.1.transformer_blocks.0.norm2.bias', 'down_blocks.2.attentions.1.transformer_blocks.0.norm2.weight', 'down_blocks.2.attentions.1.transformer_blocks.0.norm3.bias', 'down_blocks.2.attentions.1.transformer_blocks.0.norm3.weight', 'down_blocks.2.downsamplers.0.conv.bias', 'down_blocks.2.downsamplers.0.conv.weight', 'down_blocks.2.resnets.0.conv1.bias', 'down_blocks.2.resnets.0.conv1.weight', 'down_blocks.2.resnets.0.conv2.bias', 'down_blocks.2.resnets.0.conv2.weight', 'down_blocks.2.resnets.0.conv_shortcut.bias', 'down_blocks.2.resnets.0.conv_shortcut.weight', 'down_blocks.2.resnets.0.norm1.bias', 'down_blocks.2.resnets.0.norm1.weight', 'down_blocks.2.resnets.0.norm2.bias', 'down_blocks.2.resnets.0.norm2.weight', 'down_blocks.2.resnets.0.time_emb_proj.bias', 'down_blocks.2.resnets.0.time_emb_proj.weight', 'down_blocks.2.resnets.1.conv1.bias', 'down_blocks.2.resnets.1.conv1.weight', 'down_blocks.2.resnets.1.conv2.bias', 'down_blocks.2.resnets.1.conv2.weight', 'down_blocks.2.resnets.1.norm1.bias', 'down_blocks.2.resnets.1.norm1.weight', 'down_blocks.2.resnets.1.norm2.bias', 'down_blocks.2.resnets.1.norm2.weight', 'down_blocks.2.resnets.1.time_emb_proj.bias', 'down_blocks.2.resnets.1.time_emb_proj.weight', 'down_blocks.3.resnets.0.conv1.bias', 'down_blocks.3.resnets.0.conv1.weight', 'down_blocks.3.resnets.0.conv2.bias', 'down_blocks.3.resnets.0.conv2.weight', 'down_blocks.3.resnets.0.norm1.bias', 'down_blocks.3.resnets.0.norm1.weight', 'down_blocks.3.resnets.0.norm2.bias', 'down_blocks.3.resnets.0.norm2.weight', 'down_blocks.3.resnets.0.time_emb_proj.bias', 'down_blocks.3.resnets.0.time_emb_proj.weight', 'down_blocks.3.resnets.1.conv1.bias', 'down_blocks.3.resnets.1.conv1.weight', 'down_blocks.3.resnets.1.conv2.bias', 'down_blocks.3.resnets.1.conv2.weight', 'down_blocks.3.resnets.1.norm1.bias', 'down_blocks.3.resnets.1.norm1.weight', 'down_blocks.3.resnets.1.norm2.bias', 'down_blocks.3.resnets.1.norm2.weight', 'down_blocks.3.resnets.1.time_emb_proj.bias', 'down_blocks.3.resnets.1.time_emb_proj.weight', 'mid_block.attentions.0.norm.bias', 'mid_block.attentions.0.norm.weight', 'mid_block.attentions.0.proj_in.bias', 'mid_block.attentions.0.proj_in.weight', 'mid_block.attentions.0.proj_out.bias', 'mid_block.attentions.0.proj_out.weight', 'mid_block.attentions.0.transformer_blocks.0.attn1.to_k.weight', 'mid_block.attentions.0.transformer_blocks.0.attn1.to_out.0.bias', 'mid_block.attentions.0.transformer_blocks.0.attn1.to_out.0.weight', 'mid_block.attentions.0.transformer_blocks.0.attn1.to_q.weight', 'mid_block.attentions.0.transformer_blocks.0.attn1.to_v.weight', 'mid_block.attentions.0.transformer_blocks.0.attn2.to_k.weight', 'mid_block.attentions.0.transformer_blocks.0.attn2.to_out.0.bias', 'mid_block.attentions.0.transformer_blocks.0.attn2.to_out.0.weight', 'mid_block.attentions.0.transformer_blocks.0.attn2.to_q.weight', 'mid_block.attentions.0.transformer_blocks.0.attn2.to_v.weight', 'mid_block.attentions.0.transformer_blocks.0.ff.net.0.proj.bias', 'mid_block.attentions.0.transformer_blocks.0.ff.net.0.proj.weight', 'mid_block.attentions.0.transformer_blocks.0.ff.net.2.bias', 'mid_block.attentions.0.transformer_blocks.0.ff.net.2.weight', 'mid_block.attentions.0.transformer_blocks.0.norm1.bias', 'mid_block.attentions.0.transformer_blocks.0.norm1.weight', 'mid_block.attentions.0.transformer_blocks.0.norm2.bias', 'mid_block.attentions.0.transformer_blocks.0.norm2.weight', 'mid_block.attentions.0.transformer_blocks.0.norm3.bias', 'mid_block.attentions.0.transformer_blocks.0.norm3.weight', 'mid_block.resnets.0.conv1.bias', 'mid_block.resnets.0.conv1.weight', 'mid_block.resnets.0.conv2.bias', 'mid_block.resnets.0.conv2.weight', 'mid_block.resnets.0.norm1.bias', 'mid_block.resnets.0.norm1.weight', 'mid_block.resnets.0.norm2.bias', 'mid_block.resnets.0.norm2.weight', 'mid_block.resnets.0.time_emb_proj.bias', 'mid_block.resnets.0.time_emb_proj.weight', 'mid_block.resnets.1.conv1.bias', 'mid_block.resnets.1.conv1.weight', 'mid_block.resnets.1.conv2.bias', 'mid_block.resnets.1.conv2.weight', 'mid_block.resnets.1.norm1.bias', 'mid_block.resnets.1.norm1.weight', 'mid_block.resnets.1.norm2.bias', 'mid_block.resnets.1.norm2.weight', 'mid_block.resnets.1.time_emb_proj.bias', 'mid_block.resnets.1.time_emb_proj.weight', 'time_embedding.cond_proj.weight', 'time_embedding.linear_1.bias', 'time_embedding.linear_1.weight', 'time_embedding.linear_2.bias', 'time_embedding.linear_2.weight', 'up_blocks.0.resnets.0.conv1.bias', 'up_blocks.0.resnets.0.conv1.weight', 'up_blocks.0.resnets.0.conv2.bias', 'up_blocks.0.resnets.0.conv2.weight', 'up_blocks.0.resnets.0.conv_shortcut.bias', 'up_blocks.0.resnets.0.conv_shortcut.weight', 'up_blocks.0.resnets.0.norm1.bias', 'up_blocks.0.resnets.0.norm1.weight', 'up_blocks.0.resnets.0.norm2.bias', 'up_blocks.0.resnets.0.norm2.weight', 'up_blocks.0.resnets.0.time_emb_proj.bias', 'up_blocks.0.resnets.0.time_emb_proj.weight', 'up_blocks.0.resnets.1.conv1.bias', 'up_blocks.0.resnets.1.conv1.weight', 'up_blocks.0.resnets.1.conv2.bias', 'up_blocks.0.resnets.1.conv2.weight', 'up_blocks.0.resnets.1.conv_shortcut.bias', 'up_blocks.0.resnets.1.conv_shortcut.weight', 'up_blocks.0.resnets.1.norm1.bias', 'up_blocks.0.resnets.1.norm1.weight', 'up_blocks.0.resnets.1.norm2.bias', 'up_blocks.0.resnets.1.norm2.weight', 'up_blocks.0.resnets.1.time_emb_proj.bias', 'up_blocks.0.resnets.1.time_emb_proj.weight', 'up_blocks.0.resnets.2.conv1.bias', 'up_blocks.0.resnets.2.conv1.weight', 'up_blocks.0.resnets.2.conv2.bias', 'up_blocks.0.resnets.2.conv2.weight', 'up_blocks.0.resnets.2.conv_shortcut.bias', 'up_blocks.0.resnets.2.conv_shortcut.weight', 'up_blocks.0.resnets.2.norm1.bias', 'up_blocks.0.resnets.2.norm1.weight', 'up_blocks.0.resnets.2.norm2.bias', 'up_blocks.0.resnets.2.norm2.weight', 'up_blocks.0.resnets.2.time_emb_proj.bias', 'up_blocks.0.resnets.2.time_emb_proj.weight', 'up_blocks.0.upsamplers.0.conv.bias', 'up_blocks.0.upsamplers.0.conv.weight', 'up_blocks.1.attentions.0.norm.bias', 'up_blocks.1.attentions.0.norm.weight', 'up_blocks.1.attentions.0.proj_in.bias', 'up_blocks.1.attentions.0.proj_in.weight', 'up_blocks.1.attentions.0.proj_out.bias', 'up_blocks.1.attentions.0.proj_out.weight', 'up_blocks.1.attentions.0.transformer_blocks.0.attn1.to_k.weight', 'up_blocks.1.attentions.0.transformer_blocks.0.attn1.to_out.0.bias', 'up_blocks.1.attentions.0.transformer_blocks.0.attn1.to_out.0.weight', 'up_blocks.1.attentions.0.transformer_blocks.0.attn1.to_q.weight', 'up_blocks.1.attentions.0.transformer_blocks.0.attn1.to_v.weight', 'up_blocks.1.attentions.0.transformer_blocks.0.attn2.to_k.weight', 'up_blocks.1.attentions.0.transformer_blocks.0.attn2.to_out.0.bias', 'up_blocks.1.attentions.0.transformer_blocks.0.attn2.to_out.0.weight', 'up_blocks.1.attentions.0.transformer_blocks.0.attn2.to_q.weight', 'up_blocks.1.attentions.0.transformer_blocks.0.attn2.to_v.weight', 'up_blocks.1.attentions.0.transformer_blocks.0.ff.net.0.proj.bias', 'up_blocks.1.attentions.0.transformer_blocks.0.ff.net.0.proj.weight', 'up_blocks.1.attentions.0.transformer_blocks.0.ff.net.2.bias', 'up_blocks.1.attentions.0.transformer_blocks.0.ff.net.2.weight', 'up_blocks.1.attentions.0.transformer_blocks.0.norm1.bias', 'up_blocks.1.attentions.0.transformer_blocks.0.norm1.weight', 'up_blocks.1.attentions.0.transformer_blocks.0.norm2.bias', 'up_blocks.1.attentions.0.transformer_blocks.0.norm2.weight', 'up_blocks.1.attentions.0.transformer_blocks.0.norm3.bias', 'up_blocks.1.attentions.0.transformer_blocks.0.norm3.weight', 'up_blocks.1.attentions.1.norm.bias', 'up_blocks.1.attentions.1.norm.weight', 'up_blocks.1.attentions.1.proj_in.bias', 'up_blocks.1.attentions.1.proj_in.weight', 'up_blocks.1.attentions.1.proj_out.bias', 'up_blocks.1.attentions.1.proj_out.weight', 'up_blocks.1.attentions.1.transformer_blocks.0.attn1.to_k.weight', 'up_blocks.1.attentions.1.transformer_blocks.0.attn1.to_out.0.bias', 'up_blocks.1.attentions.1.transformer_blocks.0.attn1.to_out.0.weight', 'up_blocks.1.attentions.1.transformer_blocks.0.attn1.to_q.weight', 'up_blocks.1.attentions.1.transformer_blocks.0.attn1.to_v.weight', 'up_blocks.1.attentions.1.transformer_blocks.0.attn2.to_k.weight', 'up_blocks.1.attentions.1.transformer_blocks.0.attn2.to_out.0.bias', 'up_blocks.1.attentions.1.transformer_blocks.0.attn2.to_out.0.weight', 'up_blocks.1.attentions.1.transformer_blocks.0.attn2.to_q.weight', 'up_blocks.1.attentions.1.transformer_blocks.0.attn2.to_v.weight', 'up_blocks.1.attentions.1.transformer_blocks.0.ff.net.0.proj.bias', 'up_blocks.1.attentions.1.transformer_blocks.0.ff.net.0.proj.weight', 'up_blocks.1.attentions.1.transformer_blocks.0.ff.net.2.bias', 'up_blocks.1.attentions.1.transformer_blocks.0.ff.net.2.weight', 'up_blocks.1.attentions.1.transformer_blocks.0.norm1.bias', 'up_blocks.1.attentions.1.transformer_blocks.0.norm1.weight', 'up_blocks.1.attentions.1.transformer_blocks.0.norm2.bias', 'up_blocks.1.attentions.1.transformer_blocks.0.norm2.weight', 'up_blocks.1.attentions.1.transformer_blocks.0.norm3.bias', 'up_blocks.1.attentions.1.transformer_blocks.0.norm3.weight', 'up_blocks.1.attentions.2.norm.bias', 'up_blocks.1.attentions.2.norm.weight', 'up_blocks.1.attentions.2.proj_in.bias', 'up_blocks.1.attentions.2.proj_in.weight', 'up_blocks.1.attentions.2.proj_out.bias', 'up_blocks.1.attentions.2.proj_out.weight', 'up_blocks.1.attentions.2.transformer_blocks.0.attn1.to_k.weight', 'up_blocks.1.attentions.2.transformer_blocks.0.attn1.to_out.0.bias', 'up_blocks.1.attentions.2.transformer_blocks.0.attn1.to_out.0.weight', 'up_blocks.1.attentions.2.transformer_blocks.0.attn1.to_q.weight', 'up_blocks.1.attentions.2.transformer_blocks.0.attn1.to_v.weight', 'up_blocks.1.attentions.2.transformer_blocks.0.attn2.to_k.weight', 'up_blocks.1.attentions.2.transformer_blocks.0.attn2.to_out.0.bias', 'up_blocks.1.attentions.2.transformer_blocks.0.attn2.to_out.0.weight', 'up_blocks.1.attentions.2.transformer_blocks.0.attn2.to_q.weight', 'up_blocks.1.attentions.2.transformer_blocks.0.attn2.to_v.weight', 'up_blocks.1.attentions.2.transformer_blocks.0.ff.net.0.proj.bias', 'up_blocks.1.attentions.2.transformer_blocks.0.ff.net.0.proj.weight', 'up_blocks.1.attentions.2.transformer_blocks.0.ff.net.2.bias', 'up_blocks.1.attentions.2.transformer_blocks.0.ff.net.2.weight', 'up_blocks.1.attentions.2.transformer_blocks.0.norm1.bias', 'up_blocks.1.attentions.2.transformer_blocks.0.norm1.weight', 'up_blocks.1.attentions.2.transformer_blocks.0.norm2.bias', 'up_blocks.1.attentions.2.transformer_blocks.0.norm2.weight', 'up_blocks.1.attentions.2.transformer_blocks.0.norm3.bias', 'up_blocks.1.attentions.2.transformer_blocks.0.norm3.weight', 'up_blocks.1.resnets.0.conv1.bias', 'up_blocks.1.resnets.0.conv1.weight', 'up_blocks.1.resnets.0.conv2.bias', 'up_blocks.1.resnets.0.conv2.weight', 'up_blocks.1.resnets.0.conv_shortcut.bias', 'up_blocks.1.resnets.0.conv_shortcut.weight', 'up_blocks.1.resnets.0.norm1.bias', 'up_blocks.1.resnets.0.norm1.weight', 'up_blocks.1.resnets.0.norm2.bias', 'up_blocks.1.resnets.0.norm2.weight', 'up_blocks.1.resnets.0.time_emb_proj.bias', 'up_blocks.1.resnets.0.time_emb_proj.weight', 'up_blocks.1.resnets.1.conv1.bias', 'up_blocks.1.resnets.1.conv1.weight', 'up_blocks.1.resnets.1.conv2.bias', 'up_blocks.1.resnets.1.conv2.weight', 'up_blocks.1.resnets.1.conv_shortcut.bias', 'up_blocks.1.resnets.1.conv_shortcut.weight', 'up_blocks.1.resnets.1.norm1.bias', 'up_blocks.1.resnets.1.norm1.weight', 'up_blocks.1.resnets.1.norm2.bias', 'up_blocks.1.resnets.1.norm2.weight', 'up_blocks.1.resnets.1.time_emb_proj.bias', 'up_blocks.1.resnets.1.time_emb_proj.weight', 'up_blocks.1.resnets.2.conv1.bias', 'up_blocks.1.resnets.2.conv1.weight', 'up_blocks.1.resnets.2.conv2.bias', 'up_blocks.1.resnets.2.conv2.weight', 'up_blocks.1.resnets.2.conv_shortcut.bias', 'up_blocks.1.resnets.2.conv_shortcut.weight', 'up_blocks.1.resnets.2.norm1.bias', 'up_blocks.1.resnets.2.norm1.weight', 'up_blocks.1.resnets.2.norm2.bias', 'up_blocks.1.resnets.2.norm2.weight', 'up_blocks.1.resnets.2.time_emb_proj.bias', 'up_blocks.1.resnets.2.time_emb_proj.weight', 'up_blocks.1.upsamplers.0.conv.bias', 'up_blocks.1.upsamplers.0.conv.weight', 'up_blocks.2.attentions.0.norm.bias', 'up_blocks.2.attentions.0.norm.weight', 'up_blocks.2.attentions.0.proj_in.bias', 'up_blocks.2.attentions.0.proj_in.weight', 'up_blocks.2.attentions.0.proj_out.bias', 'up_blocks.2.attentions.0.proj_out.weight', 'up_blocks.2.attentions.0.transformer_blocks.0.attn1.to_k.weight', 'up_blocks.2.attentions.0.transformer_blocks.0.attn1.to_out.0.bias', 'up_blocks.2.attentions.0.transformer_blocks.0.attn1.to_out.0.weight', 'up_blocks.2.attentions.0.transformer_blocks.0.attn1.to_q.weight', 'up_blocks.2.attentions.0.transformer_blocks.0.attn1.to_v.weight', 'up_blocks.2.attentions.0.transformer_blocks.0.attn2.to_k.weight', 'up_blocks.2.attentions.0.transformer_blocks.0.attn2.to_out.0.bias', 'up_blocks.2.attentions.0.transformer_blocks.0.attn2.to_out.0.weight', 'up_blocks.2.attentions.0.transformer_blocks.0.attn2.to_q.weight', 'up_blocks.2.attentions.0.transformer_blocks.0.attn2.to_v.weight', 'up_blocks.2.attentions.0.transformer_blocks.0.ff.net.0.proj.bias', 'up_blocks.2.attentions.0.transformer_blocks.0.ff.net.0.proj.weight', 'up_blocks.2.attentions.0.transformer_blocks.0.ff.net.2.bias', 'up_blocks.2.attentions.0.transformer_blocks.0.ff.net.2.weight', 'up_blocks.2.attentions.0.transformer_blocks.0.norm1.bias', 'up_blocks.2.attentions.0.transformer_blocks.0.norm1.weight', 'up_blocks.2.attentions.0.transformer_blocks.0.norm2.bias', 'up_blocks.2.attentions.0.transformer_blocks.0.norm2.weight', 'up_blocks.2.attentions.0.transformer_blocks.0.norm3.bias', 'up_blocks.2.attentions.0.transformer_blocks.0.norm3.weight', 'up_blocks.2.attentions.1.norm.bias', 'up_blocks.2.attentions.1.norm.weight', 'up_blocks.2.attentions.1.proj_in.bias', 'up_blocks.2.attentions.1.proj_in.weight', 'up_blocks.2.attentions.1.proj_out.bias', 'up_blocks.2.attentions.1.proj_out.weight', 'up_blocks.2.attentions.1.transformer_blocks.0.attn1.to_k.weight', 'up_blocks.2.attentions.1.transformer_blocks.0.attn1.to_out.0.bias', 'up_blocks.2.attentions.1.transformer_blocks.0.attn1.to_out.0.weight', 'up_blocks.2.attentions.1.transformer_blocks.0.attn1.to_q.weight', 'up_blocks.2.attentions.1.transformer_blocks.0.attn1.to_v.weight', 'up_blocks.2.attentions.1.transformer_blocks.0.attn2.to_k.weight', 'up_blocks.2.attentions.1.transformer_blocks.0.attn2.to_out.0.bias', 'up_blocks.2.attentions.1.transformer_blocks.0.attn2.to_out.0.weight', 'up_blocks.2.attentions.1.transformer_blocks.0.attn2.to_q.weight', 'up_blocks.2.attentions.1.transformer_blocks.0.attn2.to_v.weight', 'up_blocks.2.attentions.1.transformer_blocks.0.ff.net.0.proj.bias', 'up_blocks.2.attentions.1.transformer_blocks.0.ff.net.0.proj.weight', 'up_blocks.2.attentions.1.transformer_blocks.0.ff.net.2.bias', 'up_blocks.2.attentions.1.transformer_blocks.0.ff.net.2.weight', 'up_blocks.2.attentions.1.transformer_blocks.0.norm1.bias', 'up_blocks.2.attentions.1.transformer_blocks.0.norm1.weight', 'up_blocks.2.attentions.1.transformer_blocks.0.norm2.bias', 'up_blocks.2.attentions.1.transformer_blocks.0.norm2.weight', 'up_blocks.2.attentions.1.transformer_blocks.0.norm3.bias', 'up_blocks.2.attentions.1.transformer_blocks.0.norm3.weight', 'up_blocks.2.attentions.2.norm.bias', 'up_blocks.2.attentions.2.norm.weight', 'up_blocks.2.attentions.2.proj_in.bias', 'up_blocks.2.attentions.2.proj_in.weight', 'up_blocks.2.attentions.2.proj_out.bias', 'up_blocks.2.attentions.2.proj_out.weight', 'up_blocks.2.attentions.2.transformer_blocks.0.attn1.to_k.weight', 'up_blocks.2.attentions.2.transformer_blocks.0.attn1.to_out.0.bias', 'up_blocks.2.attentions.2.transformer_blocks.0.attn1.to_out.0.weight', 'up_blocks.2.attentions.2.transformer_blocks.0.attn1.to_q.weight', 'up_blocks.2.attentions.2.transformer_blocks.0.attn1.to_v.weight', 'up_blocks.2.attentions.2.transformer_blocks.0.attn2.to_k.weight', 'up_blocks.2.attentions.2.transformer_blocks.0.attn2.to_out.0.bias', 'up_blocks.2.attentions.2.transformer_blocks.0.attn2.to_out.0.weight', 'up_blocks.2.attentions.2.transformer_blocks.0.attn2.to_q.weight', 'up_blocks.2.attentions.2.transformer_blocks.0.attn2.to_v.weight', 'up_blocks.2.attentions.2.transformer_blocks.0.ff.net.0.proj.bias', 'up_blocks.2.attentions.2.transformer_blocks.0.ff.net.0.proj.weight', 'up_blocks.2.attentions.2.transformer_blocks.0.ff.net.2.bias', 'up_blocks.2.attentions.2.transformer_blocks.0.ff.net.2.weight', 'up_blocks.2.attentions.2.transformer_blocks.0.norm1.bias', 'up_blocks.2.attentions.2.transformer_blocks.0.norm1.weight', 'up_blocks.2.attentions.2.transformer_blocks.0.norm2.bias', 'up_blocks.2.attentions.2.transformer_blocks.0.norm2.weight', 'up_blocks.2.attentions.2.transformer_blocks.0.norm3.bias', 'up_blocks.2.attentions.2.transformer_blocks.0.norm3.weight', 'up_blocks.2.resnets.0.conv1.bias', 'up_blocks.2.resnets.0.conv1.weight', 'up_blocks.2.resnets.0.conv2.bias', 'up_blocks.2.resnets.0.conv2.weight', 'up_blocks.2.resnets.0.conv_shortcut.bias', 'up_blocks.2.resnets.0.conv_shortcut.weight', 'up_blocks.2.resnets.0.norm1.bias', 'up_blocks.2.resnets.0.norm1.weight', 'up_blocks.2.resnets.0.norm2.bias', 'up_blocks.2.resnets.0.norm2.weight', 'up_blocks.2.resnets.0.time_emb_proj.bias', 'up_blocks.2.resnets.0.time_emb_proj.weight', 'up_blocks.2.resnets.1.conv1.bias', 'up_blocks.2.resnets.1.conv1.weight', 'up_blocks.2.resnets.1.conv2.bias', 'up_blocks.2.resnets.1.conv2.weight', 'up_blocks.2.resnets.1.conv_shortcut.bias', 'up_blocks.2.resnets.1.conv_shortcut.weight', 'up_blocks.2.resnets.1.norm1.bias', 'up_blocks.2.resnets.1.norm1.weight', 'up_blocks.2.resnets.1.norm2.bias', 'up_blocks.2.resnets.1.norm2.weight', 'up_blocks.2.resnets.1.time_emb_proj.bias', 'up_blocks.2.resnets.1.time_emb_proj.weight', 'up_blocks.2.resnets.2.conv1.bias', 'up_blocks.2.resnets.2.conv1.weight', 'up_blocks.2.resnets.2.conv2.bias', 'up_blocks.2.resnets.2.conv2.weight', 'up_blocks.2.resnets.2.conv_shortcut.bias', 'up_blocks.2.resnets.2.conv_shortcut.weight', 'up_blocks.2.resnets.2.norm1.bias', 'up_blocks.2.resnets.2.norm1.weight', 'up_blocks.2.resnets.2.norm2.bias', 'up_blocks.2.resnets.2.norm2.weight', 'up_blocks.2.resnets.2.time_emb_proj.bias', 'up_blocks.2.resnets.2.time_emb_proj.weight', 'up_blocks.2.upsamplers.0.conv.bias', 'up_blocks.2.upsamplers.0.conv.weight', 'up_blocks.3.attentions.0.norm.bias', 'up_blocks.3.attentions.0.norm.weight', 'up_blocks.3.attentions.0.proj_in.bias', 'up_blocks.3.attentions.0.proj_in.weight', 'up_blocks.3.attentions.0.proj_out.bias', 'up_blocks.3.attentions.0.proj_out.weight', 'up_blocks.3.attentions.0.transformer_blocks.0.attn1.to_k.weight', 'up_blocks.3.attentions.0.transformer_blocks.0.attn1.to_out.0.bias', 'up_blocks.3.attentions.0.transformer_blocks.0.attn1.to_out.0.weight', 'up_blocks.3.attentions.0.transformer_blocks.0.attn1.to_q.weight', 'up_blocks.3.attentions.0.transformer_blocks.0.attn1.to_v.weight', 'up_blocks.3.attentions.0.transformer_blocks.0.attn2.to_k.weight', 'up_blocks.3.attentions.0.transformer_blocks.0.attn2.to_out.0.bias', 'up_blocks.3.attentions.0.transformer_blocks.0.attn2.to_out.0.weight', 'up_blocks.3.attentions.0.transformer_blocks.0.attn2.to_q.weight', 'up_blocks.3.attentions.0.transformer_blocks.0.attn2.to_v.weight', 'up_blocks.3.attentions.0.transformer_blocks.0.ff.net.0.proj.bias', 'up_blocks.3.attentions.0.transformer_blocks.0.ff.net.0.proj.weight', 'up_blocks.3.attentions.0.transformer_blocks.0.ff.net.2.bias', 'up_blocks.3.attentions.0.transformer_blocks.0.ff.net.2.weight', 'up_blocks.3.attentions.0.transformer_blocks.0.norm1.bias', 'up_blocks.3.attentions.0.transformer_blocks.0.norm1.weight', 'up_blocks.3.attentions.0.transformer_blocks.0.norm2.bias', 'up_blocks.3.attentions.0.transformer_blocks.0.norm2.weight', 'up_blocks.3.attentions.0.transformer_blocks.0.norm3.bias', 'up_blocks.3.attentions.0.transformer_blocks.0.norm3.weight', 'up_blocks.3.attentions.1.norm.bias', 'up_blocks.3.attentions.1.norm.weight', 'up_blocks.3.attentions.1.proj_in.bias', 'up_blocks.3.attentions.1.proj_in.weight', 'up_blocks.3.attentions.1.proj_out.bias', 'up_blocks.3.attentions.1.proj_out.weight', 'up_blocks.3.attentions.1.transformer_blocks.0.attn1.to_k.weight', 'up_blocks.3.attentions.1.transformer_blocks.0.attn1.to_out.0.bias', 'up_blocks.3.attentions.1.transformer_blocks.0.attn1.to_out.0.weight', 'up_blocks.3.attentions.1.transformer_blocks.0.attn1.to_q.weight', 'up_blocks.3.attentions.1.transformer_blocks.0.attn1.to_v.weight', 'up_blocks.3.attentions.1.transformer_blocks.0.attn2.to_k.weight', 'up_blocks.3.attentions.1.transformer_blocks.0.attn2.to_out.0.bias', 'up_blocks.3.attentions.1.transformer_blocks.0.attn2.to_out.0.weight', 'up_blocks.3.attentions.1.transformer_blocks.0.attn2.to_q.weight', 'up_blocks.3.attentions.1.transformer_blocks.0.attn2.to_v.weight', 'up_blocks.3.attentions.1.transformer_blocks.0.ff.net.0.proj.bias', 'up_blocks.3.attentions.1.transformer_blocks.0.ff.net.0.proj.weight', 'up_blocks.3.attentions.1.transformer_blocks.0.ff.net.2.bias', 'up_blocks.3.attentions.1.transformer_blocks.0.ff.net.2.weight', 'up_blocks.3.attentions.1.transformer_blocks.0.norm1.bias', 'up_blocks.3.attentions.1.transformer_blocks.0.norm1.weight', 'up_blocks.3.attentions.1.transformer_blocks.0.norm2.bias', 'up_blocks.3.attentions.1.transformer_blocks.0.norm2.weight', 'up_blocks.3.attentions.1.transformer_blocks.0.norm3.bias', 'up_blocks.3.attentions.1.transformer_blocks.0.norm3.weight', 'up_blocks.3.attentions.2.norm.bias', 'up_blocks.3.attentions.2.norm.weight', 'up_blocks.3.attentions.2.proj_in.bias', 'up_blocks.3.attentions.2.proj_in.weight', 'up_blocks.3.attentions.2.proj_out.bias', 'up_blocks.3.attentions.2.proj_out.weight', 'up_blocks.3.attentions.2.transformer_blocks.0.attn1.to_k.weight', 'up_blocks.3.attentions.2.transformer_blocks.0.attn1.to_out.0.bias', 'up_blocks.3.attentions.2.transformer_blocks.0.attn1.to_out.0.weight', 'up_blocks.3.attentions.2.transformer_blocks.0.attn1.to_q.weight', 'up_blocks.3.attentions.2.transformer_blocks.0.attn1.to_v.weight', 'up_blocks.3.attentions.2.transformer_blocks.0.attn2.to_k.weight', 'up_blocks.3.attentions.2.transformer_blocks.0.attn2.to_out.0.bias', 'up_blocks.3.attentions.2.transformer_blocks.0.attn2.to_out.0.weight', 'up_blocks.3.attentions.2.transformer_blocks.0.attn2.to_q.weight', 'up_blocks.3.attentions.2.transformer_blocks.0.attn2.to_v.weight', 'up_blocks.3.attentions.2.transformer_blocks.0.ff.net.0.proj.bias', 'up_blocks.3.attentions.2.transformer_blocks.0.ff.net.0.proj.weight', 'up_blocks.3.attentions.2.transformer_blocks.0.ff.net.2.bias', 'up_blocks.3.attentions.2.transformer_blocks.0.ff.net.2.weight', 'up_blocks.3.attentions.2.transformer_blocks.0.norm1.bias', 'up_blocks.3.attentions.2.transformer_blocks.0.norm1.weight', 'up_blocks.3.attentions.2.transformer_blocks.0.norm2.bias', 'up_blocks.3.attentions.2.transformer_blocks.0.norm2.weight', 'up_blocks.3.attentions.2.transformer_blocks.0.norm3.bias', 'up_blocks.3.attentions.2.transformer_blocks.0.norm3.weight', 'up_blocks.3.resnets.0.conv1.bias', 'up_blocks.3.resnets.0.conv1.weight', 'up_blocks.3.resnets.0.conv2.bias', 'up_blocks.3.resnets.0.conv2.weight', 'up_blocks.3.resnets.0.conv_shortcut.bias', 'up_blocks.3.resnets.0.conv_shortcut.weight', 'up_blocks.3.resnets.0.norm1.bias', 'up_blocks.3.resnets.0.norm1.weight', 'up_blocks.3.resnets.0.norm2.bias', 'up_blocks.3.resnets.0.norm2.weight', 'up_blocks.3.resnets.0.time_emb_proj.bias', 'up_blocks.3.resnets.0.time_emb_proj.weight', 'up_blocks.3.resnets.1.conv1.bias', 'up_blocks.3.resnets.1.conv1.weight', 'up_blocks.3.resnets.1.conv2.bias', 'up_blocks.3.resnets.1.conv2.weight', 'up_blocks.3.resnets.1.conv_shortcut.bias', 'up_blocks.3.resnets.1.conv_shortcut.weight', 'up_blocks.3.resnets.1.norm1.bias', 'up_blocks.3.resnets.1.norm1.weight', 'up_blocks.3.resnets.1.norm2.bias', 'up_blocks.3.resnets.1.norm2.weight', 'up_blocks.3.resnets.1.time_emb_proj.bias', 'up_blocks.3.resnets.1.time_emb_proj.weight', 'up_blocks.3.resnets.2.conv1.bias', 'up_blocks.3.resnets.2.conv1.weight', 'up_blocks.3.resnets.2.conv2.bias', 'up_blocks.3.resnets.2.conv2.weight', 'up_blocks.3.resnets.2.conv_shortcut.bias', 'up_blocks.3.resnets.2.conv_shortcut.weight', 'up_blocks.3.resnets.2.norm1.bias', 'up_blocks.3.resnets.2.norm1.weight', 'up_blocks.3.resnets.2.norm2.bias', 'up_blocks.3.resnets.2.norm2.weight', 'up_blocks.3.resnets.2.time_emb_proj.bias', 'up_blocks.3.resnets.2.time_emb_proj.weight']

[DN6]

Hi @AbhinavGopal those keys suggest that the file is in Diffusers format and not in the original SD1.5 format that from_single_file expects.

Can you try converting the file back to the original SD1.5 using https://github.com/huggingface/diffusers/blob/main/scripts/convert_diffusers_to_original_stable_diffusion.py

[AbhinavGopal]

Thanks for the response, @DN6! Because I only wanted to convert the unet, I modified the convert_diffusers_to_original_stable_diffusion.py to this

# Script for converting a HF Diffusers saved pipeline to a Stable Diffusion checkpoint.
# *Only* converts the UNet, VAE, and Text Encoder.
# Does not convert optimizer state or any other thing.

import argparse
import os.path as osp
import re

import torch
from safetensors.torch import load_file, save_file


# =================#
# UNet Conversion #
# =================#

unet_conversion_map = [
    # (stable-diffusion, HF Diffusers)
    ("time_embed.0.weight", "time_embedding.linear_1.weight"),
    ("time_embed.0.bias", "time_embedding.linear_1.bias"),
    ("time_embed.2.weight", "time_embedding.linear_2.weight"),
    ("time_embed.2.bias", "time_embedding.linear_2.bias"),
    ("input_blocks.0.0.weight", "conv_in.weight"),
    ("input_blocks.0.0.bias", "conv_in.bias"),
    ("out.0.weight", "conv_norm_out.weight"),
    ("out.0.bias", "conv_norm_out.bias"),
    ("out.2.weight", "conv_out.weight"),
    ("out.2.bias", "conv_out.bias"),
]

unet_conversion_map_resnet = [
    # (stable-diffusion, HF Diffusers)
    ("in_layers.0", "norm1"),
    ("in_layers.2", "conv1"),
    ("out_layers.0", "norm2"),
    ("out_layers.3", "conv2"),
    ("emb_layers.1", "time_emb_proj"),
    ("skip_connection", "conv_shortcut"),
]

unet_conversion_map_layer = []
# hardcoded number of downblocks and resnets/attentions...
# would need smarter logic for other networks.
for i in range(4):
    # loop over downblocks/upblocks

    for j in range(2):
        # loop over resnets/attentions for downblocks
        hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}."
        sd_down_res_prefix = f"input_blocks.{3*i + j + 1}.0."
        unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix))

        if i < 3:
            # no attention layers in down_blocks.3
            hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}."
            sd_down_atn_prefix = f"input_blocks.{3*i + j + 1}.1."
            unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix))

    for j in range(3):
        # loop over resnets/attentions for upblocks
        hf_up_res_prefix = f"up_blocks.{i}.resnets.{j}."
        sd_up_res_prefix = f"output_blocks.{3*i + j}.0."
        unet_conversion_map_layer.append((sd_up_res_prefix, hf_up_res_prefix))

        if i > 0:
            # no attention layers in up_blocks.0
            hf_up_atn_prefix = f"up_blocks.{i}.attentions.{j}."
            sd_up_atn_prefix = f"output_blocks.{3*i + j}.1."
            unet_conversion_map_layer.append((sd_up_atn_prefix, hf_up_atn_prefix))

    if i < 3:
        # no downsample in down_blocks.3
        hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv."
        sd_downsample_prefix = f"input_blocks.{3*(i+1)}.0.op."
        unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix))

        # no upsample in up_blocks.3
        hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
        sd_upsample_prefix = f"output_blocks.{3*i + 2}.{1 if i == 0 else 2}."
        unet_conversion_map_layer.append((sd_upsample_prefix, hf_upsample_prefix))

hf_mid_atn_prefix = "mid_block.attentions.0."
sd_mid_atn_prefix = "middle_block.1."
unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix))

for j in range(2):
    hf_mid_res_prefix = f"mid_block.resnets.{j}."
    sd_mid_res_prefix = f"middle_block.{2*j}."
    unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix))


def convert_unet_state_dict(unet_state_dict):
    # buyer beware: this is a *brittle* function,
    # and correct output requires that all of these pieces interact in
    # the exact order in which I have arranged them.
    mapping = {k: k for k in unet_state_dict.keys()}
    for sd_name, hf_name in unet_conversion_map:
        mapping[hf_name] = sd_name
    for k, v in mapping.items():
        if "resnets" in k:
            for sd_part, hf_part in unet_conversion_map_resnet:
                v = v.replace(hf_part, sd_part)
            mapping[k] = v
    for k, v in mapping.items():
        for sd_part, hf_part in unet_conversion_map_layer:
            v = v.replace(hf_part, sd_part)
        mapping[k] = v
    new_state_dict = {v: unet_state_dict[k] for k, v in mapping.items()}
    return new_state_dict


# ================#
# VAE Conversion #
# ================#

vae_conversion_map = [
    # (stable-diffusion, HF Diffusers)
    ("nin_shortcut", "conv_shortcut"),
    ("norm_out", "conv_norm_out"),
    ("mid.attn_1.", "mid_block.attentions.0."),
]

for i in range(4):
    # down_blocks have two resnets
    for j in range(2):
        hf_down_prefix = f"encoder.down_blocks.{i}.resnets.{j}."
        sd_down_prefix = f"encoder.down.{i}.block.{j}."
        vae_conversion_map.append((sd_down_prefix, hf_down_prefix))

    if i < 3:
        hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0."
        sd_downsample_prefix = f"down.{i}.downsample."
        vae_conversion_map.append((sd_downsample_prefix, hf_downsample_prefix))

        hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
        sd_upsample_prefix = f"up.{3-i}.upsample."
        vae_conversion_map.append((sd_upsample_prefix, hf_upsample_prefix))

    # up_blocks have three resnets
    # also, up blocks in hf are numbered in reverse from sd
    for j in range(3):
        hf_up_prefix = f"decoder.up_blocks.{i}.resnets.{j}."
        sd_up_prefix = f"decoder.up.{3-i}.block.{j}."
        vae_conversion_map.append((sd_up_prefix, hf_up_prefix))

# this part accounts for mid blocks in both the encoder and the decoder
for i in range(2):
    hf_mid_res_prefix = f"mid_block.resnets.{i}."
    sd_mid_res_prefix = f"mid.block_{i+1}."
    vae_conversion_map.append((sd_mid_res_prefix, hf_mid_res_prefix))


vae_conversion_map_attn = [
    # (stable-diffusion, HF Diffusers)
    ("norm.", "group_norm."),
    ("q.", "query."),
    ("k.", "key."),
    ("v.", "value."),
    ("proj_out.", "proj_attn."),
]

# This is probably not the most ideal solution, but it does work.
vae_extra_conversion_map = [
    ("to_q", "q"),
    ("to_k", "k"),
    ("to_v", "v"),
    ("to_out.0", "proj_out"),
]


def reshape_weight_for_sd(w):
    # convert HF linear weights to SD conv2d weights
    if not w.ndim == 1:
        return w.reshape(*w.shape, 1, 1)
    else:
        return w


def convert_vae_state_dict(vae_state_dict):
    mapping = {k: k for k in vae_state_dict.keys()}
    for k, v in mapping.items():
        for sd_part, hf_part in vae_conversion_map:
            v = v.replace(hf_part, sd_part)
        mapping[k] = v
    for k, v in mapping.items():
        if "attentions" in k:
            for sd_part, hf_part in vae_conversion_map_attn:
                v = v.replace(hf_part, sd_part)
            mapping[k] = v
    new_state_dict = {v: vae_state_dict[k] for k, v in mapping.items()}
    weights_to_convert = ["q", "k", "v", "proj_out"]
    keys_to_rename = {}
    for k, v in new_state_dict.items():
        for weight_name in weights_to_convert:
            if f"mid.attn_1.{weight_name}.weight" in k:
                print(f"Reshaping {k} for SD format")
                new_state_dict[k] = reshape_weight_for_sd(v)
        for weight_name, real_weight_name in vae_extra_conversion_map:
            if f"mid.attn_1.{weight_name}.weight" in k or f"mid.attn_1.{weight_name}.bias" in k:
                keys_to_rename[k] = k.replace(weight_name, real_weight_name)
    for k, v in keys_to_rename.items():
        if k in new_state_dict:
            print(f"Renaming {k} to {v}")
            new_state_dict[v] = reshape_weight_for_sd(new_state_dict[k])
            del new_state_dict[k]
    return new_state_dict


# =========================#
# Text Encoder Conversion #
# =========================#


textenc_conversion_lst = [
    # (stable-diffusion, HF Diffusers)
    ("resblocks.", "text_model.encoder.layers."),
    ("ln_1", "layer_norm1"),
    ("ln_2", "layer_norm2"),
    (".c_fc.", ".fc1."),
    (".c_proj.", ".fc2."),
    (".attn", ".self_attn"),
    ("ln_final.", "transformer.text_model.final_layer_norm."),
    ("token_embedding.weight", "transformer.text_model.embeddings.token_embedding.weight"),
    ("positional_embedding", "transformer.text_model.embeddings.position_embedding.weight"),
]
protected = {re.escape(x[1]): x[0] for x in textenc_conversion_lst}
textenc_pattern = re.compile("|".join(protected.keys()))

# Ordering is from https://github.com/pytorch/pytorch/blob/master/test/cpp/api/modules.cpp
code2idx = {"q": 0, "k": 1, "v": 2}


def convert_text_enc_state_dict_v20(text_enc_dict):
    new_state_dict = {}
    capture_qkv_weight = {}
    capture_qkv_bias = {}
    for k, v in text_enc_dict.items():
        if (
            k.endswith(".self_attn.q_proj.weight")
            or k.endswith(".self_attn.k_proj.weight")
            or k.endswith(".self_attn.v_proj.weight")
        ):
            k_pre = k[: -len(".q_proj.weight")]
            k_code = k[-len("q_proj.weight")]
            if k_pre not in capture_qkv_weight:
                capture_qkv_weight[k_pre] = [None, None, None]
            capture_qkv_weight[k_pre][code2idx[k_code]] = v
            continue

        if (
            k.endswith(".self_attn.q_proj.bias")
            or k.endswith(".self_attn.k_proj.bias")
            or k.endswith(".self_attn.v_proj.bias")
        ):
            k_pre = k[: -len(".q_proj.bias")]
            k_code = k[-len("q_proj.bias")]
            if k_pre not in capture_qkv_bias:
                capture_qkv_bias[k_pre] = [None, None, None]
            capture_qkv_bias[k_pre][code2idx[k_code]] = v
            continue

        relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k)
        new_state_dict[relabelled_key] = v

    for k_pre, tensors in capture_qkv_weight.items():
        if None in tensors:
            raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
        relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
        new_state_dict[relabelled_key + ".in_proj_weight"] = torch.cat(tensors)

    for k_pre, tensors in capture_qkv_bias.items():
        if None in tensors:
            raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
        relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
        new_state_dict[relabelled_key + ".in_proj_bias"] = torch.cat(tensors)

    return new_state_dict


def convert_text_enc_state_dict(text_enc_dict):
    return text_enc_dict


if __name__ == "__main__":
    parser = argparse.ArgumentParser()

    parser.add_argument("--model_path", default=None, type=str, required=True, help="Path to the model to convert.")
    parser.add_argument("--checkpoint_path", default=None, type=str, required=True, help="Path to the output model.")
    parser.add_argument("--half", action="store_true", help="Save weights in half precision.")
    parser.add_argument(
        "--use_safetensors", action="store_true", help="Save weights use safetensors, default is ckpt."
    )

    args = parser.parse_args()

    assert args.model_path is not None, "Must provide a model path!"
    assert args.checkpoint_path is not None, "Must provide a checkpoint path!"

    # Path for safetensors
    unet_path = osp.join(args.model_path, "unet", "diffusion_pytorch_model.safetensors")
    vae_path = osp.join(args.model_path, "vae", "diffusion_pytorch_model.safetensors")
    text_enc_path = osp.join(args.model_path, "text_encoder", "model.safetensors")

    # Track which models were found
    models_found = {"unet": False, "vae": False, "text_encoder": False}
    model_state_dicts = {}

    # Try to load UNet
    if osp.exists(unet_path):
        unet_state_dict = load_file(unet_path, device="cpu")
        models_found["unet"] = True
    else:
        unet_path = osp.join(args.model_path, "unet", "diffusion_pytorch_model.bin")
        if osp.exists(unet_path):
            unet_state_dict = torch.load(unet_path, map_location="cpu")
            models_found["unet"] = True
    
    # Try to load VAE
    if osp.exists(vae_path):
        vae_state_dict = load_file(vae_path, device="cpu")
        models_found["vae"] = True
    else:
        vae_path = osp.join(args.model_path, "vae", "diffusion_pytorch_model.bin")
        if osp.exists(vae_path):
            vae_state_dict = torch.load(vae_path, map_location="cpu")
            models_found["vae"] = True
    
    # Try to load Text Encoder
    if osp.exists(text_enc_path):
        text_enc_dict = load_file(text_enc_path, device="cpu")
        models_found["text_encoder"] = True
    else:
        text_enc_path = osp.join(args.model_path, "text_encoder", "pytorch_model.bin")
        if osp.exists(text_enc_path):
            text_enc_dict = torch.load(text_enc_path, map_location="cpu")
            models_found["text_encoder"] = True

    # Process models that were found
    if models_found["unet"]:
        unet_state_dict = convert_unet_state_dict(unet_state_dict)
        unet_state_dict = {"model.diffusion_model." + k: v for k, v in unet_state_dict.items()}
        model_state_dicts["unet"] = unet_state_dict

    if models_found["vae"]:
        vae_state_dict = convert_vae_state_dict(vae_state_dict)
        vae_state_dict = {"first_stage_model." + k: v for k, v in vae_state_dict.items()}
        model_state_dicts["vae"] = vae_state_dict

    if models_found["text_encoder"]:
        # Easiest way to identify v2.0 model seems to be that the text encoder (OpenCLIP) is deeper
        is_v20_model = "text_model.encoder.layers.22.layer_norm2.bias" in text_enc_dict

        if is_v20_model:
            # Need to add the tag 'transformer' in advance so we can knock it out from the final layer-norm
            text_enc_dict = {"transformer." + k: v for k, v in text_enc_dict.items()}
            text_enc_dict = convert_text_enc_state_dict_v20(text_enc_dict)
            text_enc_dict = {"cond_stage_model.model." + k: v for k, v in text_enc_dict.items()}
        else:
            text_enc_dict = convert_text_enc_state_dict(text_enc_dict)
            text_enc_dict = {"cond_stage_model.transformer." + k: v for k, v in text_enc_dict.items()}
        
        model_state_dicts["text_encoder"] = text_enc_dict

    # Apply half precision if requested
    if args.half:
        for model_type, state_dict in model_state_dicts.items():
            model_state_dicts[model_type] = {k: v.half() for k, v in state_dict.items()}

    # Save models
    if all(models_found.values()):
        # If all models are present, combine them and save as requested
        state_dict = {**model_state_dicts.get("unet", {}), 
                     **model_state_dicts.get("vae", {}), 
                     **model_state_dicts.get("text_encoder", {})}
        
        if args.use_safetensors:
            save_file(state_dict, args.checkpoint_path)
        else:
            state_dict = {"state_dict": state_dict}
            torch.save(state_dict, args.checkpoint_path)
        print(f"Saved combined checkpoint to {args.checkpoint_path}")
    else:
        # Save individual models
        base_path = osp.splitext(args.checkpoint_path)[0]
        for model_type, state_dict in model_state_dicts.items():
            save_path = f"{base_path}_{model_type}"
            if args.use_safetensors:
                save_path += ".safetensors"
                save_file(state_dict, save_path)
            else:
                save_path += ".ckpt"
                torch.save({"state_dict": state_dict}, save_path)
            print(f"Saved {model_type} to {save_path}")

Basically, it converts the unet file into safetensors and saves it as original_unet.safetensors.

When I load it after with this:

import torch
from diffusers import UNet2DConditionModel

config = UNet2DConditionModel.load_config("SimianLuo/LCM_Dreamshaper_v7", subfolder="unet")
unet = UNet2DConditionModel.from_single_file(
"original_unet.safetensors",
config="SimianLuo/LCM_Dreamshaper_v7", # Use the repo ID string
subfolder='unet',
torch_dtype=torch.bfloat16,
use_safetensors=True
).to('cuda')

I get the following error: NotImplementedError: Cannot copy out of meta tensor; no data! Please use torch.nn.Module.to_empty() instead of torch.nn.Module.to() when moving module from meta to a different device. This tells me that the unet is maybe loaded in as empty?

[AbhinavGopal]

For clarity -- my goal is to download and convert the model into a quantized GGUF model and also compile + trace it.

[AbhinavGopal]

In order to fix the cuda issue, I used to_empty, which suppressed the error, but the output looks very noisy.
Here's my inference script:

import torch
from diffusers import AutoPipelineForImage2Image, LCMScheduler
from diffusers import UNet2DConditionModel
from PIL import Image

# Initialize Scheduler
scheduler = LCMScheduler(
    beta_start=0.00085, 
    beta_end=0.0120, 
    beta_schedule="scaled_linear", 
    prediction_type="epsilon"
)

# Load the UNet configuration
config = UNet2DConditionModel.load_config(
    "SimianLuo/LCM_Dreamshaper_v7", 
    subfolder="unet"
)

# Load the UNet from safetensors file
unet = UNet2DConditionModel.from_single_file(
    "original_unet.safetensors",
    config="SimianLuo/LCM_Dreamshaper_v7",
    subfolder='unet',
    torch_dtype=torch.bfloat16,
    use_safetensors=True
)

# Use to_empty() instead of to() for meta tensors - this creates parameter placeholders
# without allocating memory for the values
unet = unet.to_empty(device='cuda')
print("UNet loaded with to_empty()")

# Create the pipeline with the loaded UNet
pipe = AutoPipelineForImage2Image.from_pretrained(
    "SimianLuo/LCM_Dreamshaper_v7",
    torch_dtype=torch.bfloat16,
    unet=unet,
    safety_checker=None
).to('cuda')

# Move other components to CUDA with to_empty for memory efficiency
pipe.tokenizer = pipe.tokenizer
pipe.scheduler = scheduler
print("Pipeline components loaded with to_empty()")

# Process the image
init_image = Image.open("trees.jpg")
prompt = "beautiful ocean with palm trees coming out of the water, high quality, nikon, HDR"
negative_prompt = "ugly, drawing, blurry, animated, painting"

print("Starting inference...")
# Generate image
image = pipe(
    prompt=prompt, 
    image=init_image,  
    negative_prompt=negative_prompt, 
    num_images_per_prompt=1, 
    num_inference_steps=20, 
    guidance_scale=1, 
    strength=0.7
).images[0]

image.save("temp.png")
print("Image saved as temp.png") //very very noisy

[DN6]

Hmm have you tried just converting one of the dreamshaper single file checkpoints to GGUF directly?
https://huggingface.co/Lykon/DreamShaper/blob/main/DreamShaper_8_pruned.safetensors

If it's just the UNet you're interested in, you can extract with by removing the keys that contains the model.diffusion_model prefix in that file.

BTW I don't think inference with GGUF + Compile will work at the moment. We have to look into adding support for it.

[AbhinavGopal]

I needed an LCM dreamshaper checkpoint, so it needs to be this one :/

Ah, I wasn't aware of that. Is there a supported quantization method for compilation?

[DN6]

TorchAO quantization types support torch compile, we just added Quanto support to main that supports torch compile with int8 weights.

[AbhinavGopal]

Hey @DN6 , so I was looking into the conversion script, and I notice that the key "time_embedding.cond_proj.weight" in the checkpoint that I sent was being left unmodified. Even when I added this:

    ("time_embed.cond_proj.weight", "time_embedding.cond_proj.weight"),

to https://github.com/huggingface/diffusers/blob/main/scripts/convert_diffusers_to_original_stable_diffusion.py#L17, it was still noisy. I think there's possibly some bug here? I've noticed similar issues that might be related like #6848. Could you please provide some guidance?

[DN6]

@AbhinavGopal Are you still trying to convert to GGUF? Or did you switch to another quant method? (TorchAO, Quanto?)

[AbhinavGopal]

Hey @DN6! I've been using torchao, thanks for the help! Wasn't able to get GGUF up just yet.