`from_pretrained` downloads `*.bin` and `*.safetensors` weights if both are present in the repo

[apolinario]

Describe the bug

If both *.bin and *.safetensors weights exist in a repository both are downloaded even thought only the *.safetensors version are used, doubling the disk size & time to download models.

As we load *.safetensors by default, we should only download them.

Reproduction

import torch
from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16)

Logs

No response

System Info

diffusers==0.11.1

[camenduru]

maybe this is a workaround low_cpu_mem_usage=False

diffusers/src/diffusers/modeling_utils.py

Line 569 in 9be94d9

def _get_model_file(

diffusers/src/diffusers/modeling_utils.py

Line 802 in 9be94d9

def _get_model_file(

diffusers/src/diffusers/modeling_utils.py

Lines 435 to 503 in 9be94d9

	if is_safetensors_available():
	try:
	model_file = cls._get_model_file(
	pretrained_model_name_or_path,
	weights_name=SAFETENSORS_WEIGHTS_NAME,
	cache_dir=cache_dir,
	force_download=force_download,
	resume_download=resume_download,
	proxies=proxies,
	local_files_only=local_files_only,
	use_auth_token=use_auth_token,
	revision=revision,
	subfolder=subfolder,
	user_agent=user_agent,
	)
	except:
	pass
	if model_file is None:
	model_file = cls._get_model_file(
	pretrained_model_name_or_path,
	weights_name=WEIGHTS_NAME,
	cache_dir=cache_dir,
	force_download=force_download,
	resume_download=resume_download,
	proxies=proxies,
	local_files_only=local_files_only,
	use_auth_token=use_auth_token,
	revision=revision,
	subfolder=subfolder,
	user_agent=user_agent,
	)
	if low_cpu_mem_usage:
	# Instantiate model with empty weights
	with accelerate.init_empty_weights():
	config, unused_kwargs = cls.load_config(
	config_path,
	cache_dir=cache_dir,
	return_unused_kwargs=True,
	force_download=force_download,
	resume_download=resume_download,
	proxies=proxies,
	local_files_only=local_files_only,
	use_auth_token=use_auth_token,
	revision=revision,
	subfolder=subfolder,
	device_map=device_map,
	**kwargs,
	)
	model = cls.from_config(config, **unused_kwargs)
	# if device_map is Non,e load the state dict on move the params from meta device to the cpu
	if device_map is None:
	param_device = "cpu"
	state_dict = load_state_dict(model_file)
	# move the parms from meta device to cpu
	for param_name, param in state_dict.items():
	set_module_tensor_to_device(model, param_name, param_device, value=param)
	else: # else let accelerate handle loading and dispatching.
	# Load weights and dispatch according to the device_map
	# by deafult the device_map is None and the weights are loaded on the CPU
	accelerate.load_checkpoint_and_dispatch(model, model_file, device_map)
	loading_info = {
	"missing_keys": [],
	"unexpected_keys": [],
	"mismatched_keys": [],
	"error_msgs": [],
	}

diffusers/src/diffusers/modeling_utils.py

Line 601 in 9be94d9

model_file = hf_hub_download(

diffusers/src/diffusers/modeling_utils.py

Line 834 in 9be94d9

model_file = hf_hub_download(

[camenduru]

also diffusers generating new snapshot again and again snapshots pointing to same thing but idk is this a bug? or is it downloading when new version in the main repo?

[julien-c]

@camenduru do you use Developer mode on Windows? If yes, those snapshots would be symlink'ed to the actual files (i.E. not duplicated on disk space)

[camenduru]

Hi @julien-c 👋 yes snapshots pointing to same thing it is not a bug but I think this is a bug 🐜

[pcuenca]

@apolinario I'm not able to reproduce this using diffusers @ main. If I delete my cache directory and then load the stable diffusion pipeline for model stabilityai/stable-diffusion-2-1, this is what ends up being downloaded:

pedro@deep-hack:diffusers$ tree models--stabilityai--stable-diffusion-2-1/
models--stabilityai--stable-diffusion-2-1/
├── blobs
│   ├── 1238522277c48923ff2751e238f2742c562e45643f3d50cc93d163cb30638b0c
│   ├── 469be27c5c010538f845f518c4f5e8574c78f7c8
│   ├── 5294955ff7801083f720b34b55d0f1f51313c5c5
│   ├── 536b82b4e3c62c4898b4ac8725bc514f2a98f5de
│   ├── 76e821f1b6f0a9709293c3b6b51ed90980b3166b
│   ├── 9b1458658e8651398962171a8c5c56c5c0bd5aea
│   ├── 9c60528fdcb99a7caf834426a94ea13c56cf422b
│   ├── a1d993488569e928462932c8c38a0760b874d166399b14414135bd9c42df5815
│   ├── ae0c5be6f35217e51c4c000fd325d8de0294e99c
│   ├── cce6febb0b6d876ee5eb24af35e27e764eb4f9b1d0b7c026c8c3333d4cfc916c
│   ├── cefd6c7732bbda467ed1ec3e33cfa332a5fe32cf
│   ├── f4fe219b936c0e171504b4bba0c33c7bef6ea211
│   └── f97af6a6a8235236b1346312f328569ce2d70f81
├── refs
│   └── main
└── snapshots
    └── 36a01dc742066de2e8c91e7cf0b8f6b53ef53da1
        ├── feature_extractor
        │   └── preprocessor_config.json -> ../../../blobs/5294955ff7801083f720b34b55d0f1f51313c5c5
        ├── model_index.json -> ../../blobs/cefd6c7732bbda467ed1ec3e33cfa332a5fe32cf
        ├── scheduler
        │   └── scheduler_config.json -> ../../../blobs/536b82b4e3c62c4898b4ac8725bc514f2a98f5de
        ├── text_encoder
        │   ├── config.json -> ../../../blobs/9c60528fdcb99a7caf834426a94ea13c56cf422b
        │   └── model.safetensors -> ../../../blobs/cce6febb0b6d876ee5eb24af35e27e764eb4f9b1d0b7c026c8c3333d4cfc916c
        ├── tokenizer
        │   ├── merges.txt -> ../../../blobs/76e821f1b6f0a9709293c3b6b51ed90980b3166b
        │   ├── special_tokens_map.json -> ../../../blobs/ae0c5be6f35217e51c4c000fd325d8de0294e99c
        │   ├── tokenizer_config.json -> ../../../blobs/f4fe219b936c0e171504b4bba0c33c7bef6ea211
        │   └── vocab.json -> ../../../blobs/469be27c5c010538f845f518c4f5e8574c78f7c8
        ├── unet
        │   ├── config.json -> ../../../blobs/9b1458658e8651398962171a8c5c56c5c0bd5aea
        │   └── diffusion_pytorch_model.safetensors -> ../../../blobs/1238522277c48923ff2751e238f2742c562e45643f3d50cc93d163cb30638b0c
        └── vae
            ├── config.json -> ../../../blobs/f97af6a6a8235236b1346312f328569ce2d70f81
            └── diffusion_pytorch_model.safetensors -> ../../../blobs/a1d993488569e928462932c8c38a0760b874d166399b14414135bd9c42df5815

10 directories, 27 files

Edit: tested with diffusers 0.11.1 too.

[pcuenca]

Ok, I've been able to reproduce it when a repo contains safetensors versions of just some files, but not all. This is the case for @camenduru's example.