Example of loading from_single_file with local_files_only=True
#6836
-
|
Hello, I am struggling to create a pipeline that would load a safetensors file, using
I guess I am looking for something like this: from diffusers import StableDiffusionXLPipeline
import torch
pipeline = StableDiffusionXLPipeline.from_single_file(
'/models/albedobond/albedobase-xl-v2.0.safetensors',
text_encoder=...,
text_encoder_2=...,
tokenizer=...,
tokenizer_2=...,
torch_dtype=torch.float16,
local_files_only=True,
use_safetensors=True,
add_watermarker=False
)I could not figure out how to supply the text encoders and tokenizers by using the local model files. How I can I load the pipeline? Also once the pipeline is loaded, how can I switch between schedulers without downloading anything from the hub? Thanks |
Beta Was this translation helpful? Give feedback.
Replies: 4 comments 8 replies
-
|
Hi, you don't have to pass the text_encoders or the tokenizers, those are managed inside but you'll need to save the config files in same paths inside the root directory from where you're running your code, so you'll need this file structure:
Also if you don't have internet access, you'll need to provide the original_config_file file for the model, so it would be something like this: pipeline = StableDiffusionXLPipeline.from_single_file(
'/models/albedobond/albedobase-xl-v2.0.safetensors',
original_config_file='path/to/sdxl/config/sd_xl_base.yaml',
torch_dtype=torch.float16,
local_files_only=True,
use_safetensors=True,
add_watermarker=False
)diffusers doesn't download anything from the hub when changing schedulers, you can do it like this: pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config)
pipeline.scheduler.config.use_karras_sigmas = True |
Beta Was this translation helpful? Give feedback.
-
|
Thank you @asomoza I'll try what you describe! |
Beta Was this translation helpful? Give feedback.
-
|
Hello @asomoza I was able to run the code offline as per your suggestion! Thank you very much. One thing though, I did not bother specifying any |
Beta Was this translation helpful? Give feedback.
-
|
Glad it worked @alexisrolland You have to specify the original config file only if you don't have internet, diffusers downloads the file with a requests not the hugging face hub, so even if you don't specify the config file it should work and you won't see the file locally anywhere. For example, if I block python with my firewall and don't specify the config_file I get this error: you can see this in the code too: diffusers/src/diffusers/loaders/single_file_utils.py Lines 297 to 310 in fbdf26b |
Beta Was this translation helpful? Give feedback.
-
|
Remember to mark a comment as an answer in case it solved your queries :) |
Beta Was this translation helpful? Give feedback.
-
|
hello @asomoza It's me again... I am wondering if it's possible to specify a custom folder to load the text encoders instead of using Example
I have tried something with pipeline = StableDiffusionXLPipeline.from_single_file(
'/models/albedobond/albedobase-xl-v2.0.safetensors',
torch_dtype=torch.float16,
cache_dir='/models',
local_files_only=True,
use_safetensors=True
)
|
Beta Was this translation helpful? Give feedback.
-
|
Hi, you can load them separately and then pass them to the pipeline from transformers import CLIPTextModel, CLIPTextModelWithProjection
text_encoder = CLIPTextModel.from_pretrained(
"path/to/text_encoder",
variant="fp16",
torch_dtype=torch.float16,
local_files_only=True,
)
text_encoder_2 = CLIPTextModelWithProjection.from_pretrained(
"path/to/text_encoder_2",
variant="fp16",
torch_dtype=torch.float16,
local_files_only=True,
)
pipeline = StableDiffusionXLPipeline.from_single_file(
'/models/albedobond/albedobase-xl-v2.0.safetensors',
original_config_file='path/to/sdxl/config/sd_xl_base.yaml',
torch_dtype=torch.float16,
local_files_only=True,
use_safetensors=True,
add_watermarker=False,
text_encoder=text_encoder,
text_encoder_2=text_encoder_2,
)does this work for you? |
Beta Was this translation helpful? Give feedback.
-
|
It does, thank you! |
Beta Was this translation helpful? Give feedback.
-
|
@asomoza but I got this error: |
Beta Was this translation helpful? Give feedback.
-
|
Hi, that error usually happens when you're trying to load a file that is not in the correct format. The error in your case is being thrown trying to load the text encoder, are you sure you have the text encoder in |
Beta Was this translation helpful? Give feedback.
-
|
Thanks to @asomoza , above error have been sloved(I did not download these lfs files), but I got another error: Below is my model file tree: And throw your above example, I try to use "openai/clip-vit-large-patch14" and "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k" as my text_encoder. But got another error: I try to remove "tf_model.h5" and "flax_model.msgpack", I try to commented code “fp16”, got result: My project file tree: Anything could you help me? Thanks! |
Beta Was this translation helpful? Give feedback.
-
|
Thanks to @asomoza , finally I succeed run this example, by change path to "stable-diffusion-xl-base-1.0/text_encoder". code: project file tree: sdxl model file tree: succeed result: |
Beta Was this translation helpful? Give feedback.
Hi, you don't have to pass the text_encoders or the tokenizers, those are managed inside but you'll need to save the config files in same paths inside the root directory from where you're running your code, so you'll need this file structure:
project_root/laion/CLIP-ViT-bigG-14-laion2B-39B-b160kproject_root/openai/clip-vit-large-patch14Also if you don't have internet access, you'll need to provide the original_config_file file for the model, so it would be something like this: