ValueError: CLIPVisionModelWithProjection does not support `device_map='auto'`. To implement support, the model class needs to implement the `_no_split_modules` attribute.

[Sosycs]

Hello, I am trying to run the code on the colab provided. I have not change anything in the code yet. after I ran the part:

from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "DeepFloyd/IF-I-XL-v1.0",
    text_encoder=text_encoder, # pass the previously instantiated 8bit text encoder
    unet=None,
    device_map="auto"
)

I got the following error:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
[<ipython-input-8-55e2717ce150>](https://localhost:8080/#) in <cell line: 3>()
      1 from diffusers import DiffusionPipeline
      2 
----> 3 pipe = DiffusionPipeline.from_pretrained(
      4     "DeepFloyd/IF-I-XL-v1.0",
      5     text_encoder=text_encoder, # pass the previously instantiated 8bit text encoder

3 frames
[/usr/local/lib/python3.10/dist-packages/transformers/modeling_utils.py](https://localhost:8080/#) in _get_no_split_modules(self, device_map)
   1688                 if isinstance(module, PreTrainedModel):
   1689                     if module._no_split_modules is None:
-> 1690                         raise ValueError(
   1691                             f"{module.__class__.__name__} does not support `device_map='{device_map}'`. To implement support, the model "
   1692                             "class needs to implement the `_no_split_modules` attribute."

ValueError: CLIPVisionModelWithProjection does not support `device_map='auto'`. To implement support, the model class needs to implement the `_no_split_modules` attribute.

and it only worked after I removed the device_map="auto".

[robgon-art]

I'm seeing the same thing...

[hda-xian]

I got the same error.

[shanghongsim]

I got the same error. May I know when it will be resolved?

[hda-xian]

这是来自QQ邮箱的假期自动回复邮件。你好，我最近正在休假中，无法亲自回复你的邮件。我将在假期结束后，尽快给你回复。

[hda-xian]

这是来自QQ邮箱的假期自动回复邮件。你好，我最近正在休假中，无法亲自回复你的邮件。我将在假期结束后，尽快给你回复。

[John-Franklin]

If anyone else runs into this, I fit in pipe = DiffusionPipeline.from_pretrained( "DeepFloyd/IF-I-XL-v1.0", text_encoder=text_encoder, # pass the previously instantiated 8bit text encoder unet=None, max_memory={0:"15GB","cpu": "13GB"} ) for that block and it worked. I also ran into a followup bug where I needed to add .to("cuda") to the pipelines as float16 is only compatible with GPU but it points to both GPU and CPU causing failures. On step 1.4, put: pipe = DiffusionPipeline.from_pretrained( "DeepFloyd/IF-I-XL-v1.0", text_encoder=None, variant="fp16", torch_dtype=torch.float16 ).to("cuda") and do similarly for other steps; you should be able to generate an image on colab.