Add Stable Diffusion ControlNet support

[JingyaHuang]

What does this PR do?

Fixes #575

• Add ControlNet export support

• Regular SD

optimum-cli export neuron -m runwayml/stable-diffusion-v1-5 --task stable-diffusion --batch_size 1 --height 512 --width 512 --controlnet_ids lllyasviel/sd-controlnet-canny --num_images_per_prompt 1 sd_neuron_controlnet/

• Tiny test

optimum-cli export neuron -m hf-internal-testing/tiny-stable-diffusion-torch --task stable-diffusion --batch_size 1 --height 64 --width 64 --controlnet_ids hf-internal-testing/tiny-controlnet --num_images_per_prompt 1 sd_neuron_tiny_controlnet/

• Add ControlNet pipeline for stable diffusion

• Compilation

from optimum.neuron import NeuronStableDiffusionControlNetPipeline

model_id = "runwayml/stable-diffusion-v1-5"
controlnet_id = "lllyasviel/sd-controlnet-canny"
save_directory = "sd_neuron_controlnet"

# [Neuron] pipeline
input_shapes = {"batch_size": 1, "height": 512, "width": 512, "num_images_per_prompt": 1}
compiler_args = {"auto_cast": "matmul", "auto_cast_type": "bf16"}
pipe = NeuronStableDiffusionControlNetPipeline.from_pretrained(
    model_id,
    controlnet_ids=controlnet_id,
    export=True,
    **input_shapes,
    **compiler_args,
)
pipe.save_pretrained(save_directory)

• Inference

import cv2
import numpy as np
from diffusers import UniPCMultistepScheduler
from diffusers.utils import load_image, make_image_grid
from PIL import Image

from optimum.neuron import NeuronStableDiffusionControlNetPipeline


# prepare canny image
original_image = load_image(
    "https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png"
)

image = np.array(original_image)

low_threshold = 100
high_threshold = 200

image = cv2.Canny(image, low_threshold, high_threshold)
image = image[:, :, None]
image = np.concatenate([image, image, image], axis=2)
canny_image = Image.fromarray(image)

# [Neuron] pipeline
save_directory = "sd_neuron_controlnet"
pipe = NeuronStableDiffusionControlNetPipeline.from_pretrained(save_directory)

pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
output = pipe("the mona lisa", image=canny_image).images[0]
compare = make_image_grid([original_image, canny_image, output], rows=1, cols=3)
compare.save("compare.png")

• Tests
• Documentation

Next Steps

• Add ControlNet pipeline for sdxl.
• Mutil ControlNets might not be well supported need to be further validated.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[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.

[Suprhimp]

what happen? I hope it will be merge it soon, aslo sdxl control model too :)

[dacorvo]

Thanks for the pull-request. Looks good to me, but I realize I missed the input flattening from previous SD pipelines pull-requests: could you explain a little bit what happens and when ?
Also, perhaps consider raising an exception when someone tries to call the SDXL ControlNet pipeline that is present but not implemented.

[dacorvo]

If you push your pull-request back, consider cherry-picking this commit from my branch to fix the TGI docker build.

[JingyaHuang]

@dacorvo for the tracing, the compiler only accepts tensors, but not a list or a tuple of tensors which could be the case in transformers. So we flatten (actually we create directly non-nested dummy) inputs during the tracing, and during the inference runtime, we need to flatten inputs generated by the preprocessor (or the output of another model in the pipe like the case of stable diffusion) before feeding it into the compiled model.

[michaelbenayoun]

Left a few nits, otherwise lgtm from what I can understand.