use_safetensors Fails on Windows When There Are Alternative bin Files

[NullSenseStudio]

Describe the bug

When a model contains safetensors and bin weights diffusers will only download the bin weights. An error is raised when use_safetensors=True:

OSError: Could not found the necessary `safetensors` weights in {'vae/diffusion_pytorch_model.fp16.bin', 'text_encoder/pytorch_model.fp16.bin', 'unet/diffusion_pytorch_model.fp16.safetensors', 'text_encoder/model.fp16.safetensors', 'vae/diffusion_pytorch_model.fp16.safetensors', 'unet/diffusion_pytorch_model.fp16.bin', 'safety_checker/pytorch_model.fp16.bin', 'safety_checker/model.fp16.safetensors'} (variant=fp16)

With a warning before it like:
vae\diffusion_pytorch_model.fp16.safetensors not found

This is happening due to a windows separator being inserted where a unix separator is expected.

diffusers/src/diffusers/pipelines/pipeline_utils.py

Line 177 in a10107f

expected_sf_filename = os.path.join(path, filename)

Reproduction

from diffusers import DiffusionPipeline
import torch

pipe = DiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16, variant="fp16", use_safetensors=True
)

A model like SDXL 1.0 doesn't have this issue due to not having bin weights to compare against.

from diffusers import DiffusionPipeline
import torch

pipe = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16, variant="fp16", use_safetensors=True
)

Logs

No response

System Info

platform: Windows 11
python: 3.10.7
diffusers: 0.19.3

Who can help?

@patrickvonplaten

[sayakpaul]

Thanks for reporting this!

@DN6 could you look into it?

[DN6]

On it 👍🏽

[patrickvonplaten]

Gentle ping here @DN6