[Scheduler] fix: EDM schedulers when using the exp sigma schedule.

[sayakpaul]

What does this PR do?

Otherwise the following would fail:

import torch
from huggingface_hub import hf_hub_download
from diffusers import StableDiffusionXLInstructPix2PixPipeline, EDMEulerScheduler
from diffusers.utils import load_image

edit_file = hf_hub_download(repo_id="stabilityai/cosxl", filename="cosxl_edit.safetensors")

pipe_edit = StableDiffusionXLInstructPix2PixPipeline.from_single_file(edit_file, num_in_channels=8, is_cosxl_edit=True)

pipe_edit.scheduler = EDMEulerScheduler(
    sigma_min=0.002, sigma_max=120.0, sigma_data=1.0, prediction_type="v_prediction", sigma_schedule="exponential"
)
pipe_edit.to("cuda")

prompt = "make it a cloudy day"
image = load_image("https://huggingface.co/datasets/multimodalart/genai-book-images/resolve/main/mountain.png")
pipe_edit(prompt=prompt, image=image).images[0]

@apolinario FYI.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[yiyixuxu]

we make the ramp variable here and it is a numpy

ramp = np.linspace(0, 1, self.num_inference_steps)

[sayakpaul]

Sorry, I didn't get your point.

We compute ramp using torch.linspace() here:

diffusers/src/diffusers/schedulers/scheduling_edm_euler.py

Line 103 in 4d633bf

ramp = torch.linspace(0, 1, num_train_timesteps)

And then we compute it with np.linspace() here:

diffusers/src/diffusers/schedulers/scheduling_edm_euler.py

Line 213 in 4d633bf

ramp = np.linspace(0, 1, self.num_inference_steps)

[yiyixuxu]

we make ramp we can decide what they are - so there is no need to have the if here :)

[sayakpaul]

@yiyixuxu LMK if the recent changes work for you.

[yiyixuxu]

test still fails though

let's update it here to be a torch tensor and remove this line

so it is consistent between set_timesteps and __init__ (I think it's causing the error) and also, we don't need to convert it to numpy and back to torch