Add StableDiffusion Dreambooth example

[0xdevalias]

Currently there are StableDiffusion examples here:

• https://github.com/facebookincubator/AITemplate/tree/main/examples/05_stable_diffusion

But it would be awesome if there was also examples of using this for Dreambooth:

• https://dreambooth.github.io/
• https://github.com/huggingface/diffusers/tree/main/examples/dreambooth

See also:

• Stable Diffusion Meta AITemplate with >= 200% performance increase AUTOMATIC1111/stable-diffusion-webui#1625
• Couldn't using META's new inference engine increase Fast Stable's speed even more? TheLastBen/fast-stable-diffusion#73
• Implementing AITemplate ShivamShrirao/diffusers#32

And similar end-goal but these aren't based on AITemplate:

• [FEATURE]: Integrate StableDiffusion Dreambooth with ColossalAI hpcaitech/ColossalAI#1935
• Community Integration: Making AIGC cheaper, faster, and more efficient. huggingface/diffusers#1212
• [Feature Request]: colossalai integration AUTOMATIC1111/stable-diffusion-webui#4606
• Integrate with ColossalAI for memory/speed efficiency and improvements d8ahazard/sd_dreambooth_extension#174
• Integrate StableDiffusion Dreambooth with ColossalAI TheLastBen/fast-stable-diffusion#493
• Integrate StableDiffusion Dreambooth with ColossalAI ShivamShrirao/diffusers#129

[0xdevalias]

AITemplate is for inference. I don't think it would provide any speedup during training.
EDIT: nevermind. You can still accelerate the frozen parts of the model.

Originally posted by @matteoserva in ShivamShrirao/diffusers#32 (comment)

It is just for inference so won't be helpful in training. I also tested it, it's good for inference but also takes a really long time to compile.

Originally posted by @ShivamShrirao in ShivamShrirao/diffusers#32 (comment)

[chenyang78]

It is just for inference so won't be helpful in training. I also tested it, it's good for inference but also takes a really long time to compile.

Just FYI - The compilation time with the latest open-source version has been improved a lot from our first release. In our experiences, it can be 4X faster for the models where computation-intensive ops are mostly GEMM-family ops. We've made similar improvement for Conv ops in our internal version, which will be sync-ed to the open-source repo later. Stay tuned. Thanks.