The huggingface repo need to be fixed for Sana 2K and 4K models

[nitinmukesh]

Describe the bug

Hello @lawrence-cj ,

I am using Sana using diffusers. The issue is applicable for both these repos and maybe for 512/1024 but not tested.

if inference_type == "Sana 4K":
    model_path = "Efficient-Large-Model/Sana_1600M_4Kpx_BF16_diffusers"
else:
    model_path = "Efficient-Large-Model/Sana_1600M_2Kpx_BF16_diffusers"

pipe_sana = SanaPipeline.from_pretrained(
    pretrained_model_name_or_path=model_path,
    variant="bf16",
    torch_dtype=torch.bfloat16,
    use_safetensors=True,
)  

When I specify bf16 and use_safetensors it should only download bf16 models and not 32 bit. It is working fine for text_encoder and vae but not for transformer.

C:\Users\nitin\.cache\huggingface\hub\models--Efficient-Large-Model--Sana_1600M_2Kpx_BF16_diffusers\snapshots\c096bbd4f6da0daf181f4fbce5e7505051b8c75c>tree /F
Folder PATH listing for volume Windows-SSD
Volume serial number is CE9F-A6AE
C:.
│   model_index.json
│
├───scheduler
│       scheduler_config.json
│
├───text_encoder
│       config.json
│       model.bf16-00001-of-00002.safetensors
│       model.bf16-00002-of-00002.safetensors
│       model.safetensors.index.bf16.json
│
├───tokenizer
│       special_tokens_map.json
│       tokenizer.json
│       tokenizer.model
│       tokenizer_config.json
│
├───transformer
│       config.json
│       diffusion_pytorch_model-00001-of-00002.safetensors
│       diffusion_pytorch_model-00002-of-00002.safetensors
│       diffusion_pytorch_model.bf16.safetensors
│       diffusion_pytorch_model.safetensors.index.json
│
└───vae
        config.json
        diffusion_pytorch_model.bf16.safetensors

Reproduction

from diffusers import SanaPipeline
model_path = "Efficient-Large-Model/Sana_1600M_2Kpx_BF16_diffusers"
pipe_sana = SanaPipeline.from_pretrained(
        pretrained_model_name_or_path=model_path,
        variant="bf16",
        torch_dtype=torch.bfloat16,
        use_safetensors=True,
    )
pipe_sana.to("cuda")
pipe_sana.vae.to(torch.bfloat16)
pipe_sana.text_encoder.to(torch.bfloat16)

Logs

Message during inference


A mixture of bf16 and non-bf16 filenames will be loaded.
Loaded bf16 filenames:
[transformer/diffusion_pytorch_model.bf16.safetensors, text_encoder/model.bf16-00001-of-00002.safetensors, text_encoder/model.bf16-00002-of-00002.safetensors, vae/diffusion_pytorch_model.bf16.safetensors]
Loaded non-bf16 filenames:
[transformer/diffusion_pytorch_model-00001-of-00002.safetensors, transformer/diffusion_pytorch_model-00002-of-00002.safetensors
If this behavior is not expected, please check your folder structure.

System Info

Not needed as this is huggingface repo setup issue

Who can help?

@lawrence-cj

[DN6]

This is a bug in how we're fetching variants. Will take a look.