Implementation of diff. merge
#2
by
alfredplpl
- opened
Great work!
According to you said, I implemented the diff. merge:
import torch
from diffusers import StableDiffusionXLPipeline
from safetensors.torch import load_file, save_file
from huggingface_hub import hf_hub_download
w=1.3
ssd2sdxl={
# your gist
}
file_name=hf_hub_download(repo_id="alfredplpl/emi", subfolder="unet", filename="diffusion_pytorch_model.safetensors")
emi = load_file(file_name)
file_name=hf_hub_download(repo_id="stabilityai/stable-diffusion-xl-base-1.0", subfolder="unet", filename="diffusion_pytorch_model.safetensors")
sdxl = load_file(file_name)
file_name=hf_hub_download(repo_id="segmind/SSD-1B", subfolder="unet", filename="diffusion_pytorch_model.fp16.safetensors")
ssd = load_file(file_name)
diff=sdxl
for k,v in diff.items():
diff[k]=emi[k]-v
for ssd_key,sdxl_key in ssd2sdxl.items():
for to_x in [".to_q",".to_k",".to_v",".to_out.0"]:
diff[sdxl_key+to_x+".weight"]=ssd[ssd_key+to_x+".weight"]+w*diff[sdxl_key+to_x+".weight"]
emi_mid=StableDiffusionXLPipeline.from_pretrained("alfredplpl/emi",torch_dtype=torch.float32)
emi_mid.save_pretrained('/mnt/NVM/demi_mid' )
save_file(diff, '/mnt/NVM/demi_mid/unet/diffusion_pytorch_model.safetensors' )
Is it right?
Thanks in advance.
Merge code is bellow.
Note that transformer is not only q,k,v,o.
import json
with open("ssd2sdxl.json","r") as f:
ssd2sdxl = json.load(f)
new_dic = {}
def replace_string_using_dict(s, replacement_dict):
for key, value in replacement_dict.items():
s = s.replace(key, value)
return s
for key in ssd:
sdxl_key = replace_string_using_dict(key, ssd2sdxl)
new_dic[key] = ssd[key] + 1.3 * (neko[sdxl_key] - sdxl[sdxl_key])
May I ask how good the merges are without any training?
Amazing!
I could merge the models!
for key in ssd:
sdxl_key = replace_string_using_dict(key, ssd2sdxl)
new_dic[key] = ssd[key] + 2.0 * (emi[sdxl_key] - sdxl[sdxl_key])
emi_mid=StableDiffusionXLPipeline.from_pretrained("segmind/SSD-1B", torch_dtype=torch.float16, use_safetensors=True, variant="fp16")
emi_mid.save_pretrained('/mnt/NVM/demi_mid' )
save_file(new_dic, '/mnt/NVM/demi_mid/unet/diffusion_pytorch_model.safetensors' )
It is very fast.
I will continue to do distillation it or to fine tune it.
@Icar Your turn
It is not all that bad apparently with just a direct merge. Finetuning / distillation should be faster with this model.
FYI: I success to finetune it.
I will close this discussion.
alfredplpl
changed discussion status to
closed