When I want to run the inference after training dreambooth, I report an error

[1235555555shan]

Model/Pipeline/Scheduler description

pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to("cuda:3")

prompt = "A photo of a beautiful Asian sks girl"
image = pipe(prompt, num_inference_steps=50, guidance_scale=7.5).images[0]

image.save("dog-bucket.png")

Open source status

• The model implementation is available
• The model weights are available (Only relevant if addition is not a scheduler).

Provide useful links for the implementation

[patrickvonplaten]

cc @williamberman

[williamberman]

Ah good catch, so this is because accelerator.save_state/load_state is just used for saving and loading the training components passed to accelerator.prepare

https://huggingface.co/docs/accelerate/v0.16.0/en/package_reference/accelerator#accelerate.Accelerator.save_state

So the cli message here might be a bit confusing and will require a bit extra docs on how to use intermediate checkpoints as pipelines.

"Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
            " checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
            " training using `--resume_from_checkpoint`."

Let me test out on my own machine and then will circle back

[williamberman]

@1235555555shan We actually already have docs for this here https://huggingface.co/docs/diffusers/main/en/training/dreambooth#performing-inference-using-a-saved-checkpoint :)

They were in the README but I'm going to also add them to the cli help message

[williamberman]

@1235555555shan I'm going to close the issue given we merged the test and better docs. Feel free to re-open if I missed something or it's not clear. Thanks for the issue report!