[SD3] CFG Cutoff fix and official callback

[asomoza]

What does this PR do?

To do a CFG cutoff with the SD3 pipeline we need to also expose the pooled_prompt_embeds and also change the shape of them.

This PR also adds an official callback to test this.

How to test

import torch

from diffusers import StableDiffusion3Pipeline
from diffusers.callbacks import SD3CFGCutoffCallback


torch_dtype = torch.bfloat16

pipe = StableDiffusion3Pipeline.from_pretrained("stabilityai/stable-diffusion-3.5-medium", torch_dtype=torch_dtype)
pipe.enable_model_cpu_offload()

cfg_callback = SD3CFGCutoffCallback(cutoff_step_ratio=0.3)

prompt = "A capybara in a field of flowers"

image = pipe(
    prompt,
    num_inference_steps=28,
    callback_on_step_end=cfg_callback,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]

image.save("capybara_cfg_cutoff.png")

No CFG cutoff	CFG cutoff

Who can review?

@yiyixuxu @cubiq

[yiyixuxu]

thanks!

[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.

[asomoza]

failing test is not related.