JAX Integration

[patrickvonplaten]

JAX Integration

This issue will be used as a tracker to integrate Stable Diffusion in JAX natively to diffusers. This will enable many cool use cases noteably running stable diffusion on a google colab.

General design:

We will make loosen the forced PyTorch dependency and instead force the user to either install PyTorch or JAX. Then we will mirror the following "base" classes to be JAX compatible:

ModelMixin: patil-suraj/stable-diffusion-jax#10 we should add a FlaxModelMixin class here.
FlaxDiffusionPipeline:

diffusers/src/diffusers/pipeline_utils.py

Line 76 in 25a51b6

class DiffusionPipeline(ConfigMixin):

we should add a FlaxDiffusionPipeline here.

Note: ModelMixin should be made state-less by default. E.g. weights will not be saved. Also contrary to transformers should we maybe only work with flax.linen.Module classes here @patil-suraj - I don't really think we need the UNetConditionModel and UNetConditionModule design here - we could just go for class UNetConditionModel(nn.Module): here and make sure everything stays stateless no?

TODO:

• 1. Make diffusers framework independent. This will require some general changes to setup.py and our automation tools
• 2. Add FlaxModelMixin: Implement FlaxModelMixin #493 Here we can take a lot from https://github.com/patil-suraj/stable-diffusion-jax/pull/10/files but I'm not sure we should follow the transformers design here 1-to-1 . Will also ask some google-folks here
• 3. Add all the modeling code under unet_2d_condition_flax.py ...
• 4. Add automatic conversion PT <=> Flax script. #478
• 5. Add PNDM scheduler under scheduling_pndm_flax.py
• 6. Tests
• 7. Create pipeline and also FlaxDiffusionPipeline

Happy to take over 1. and finish today and then look into 4. once 3. is done.

@mishig25 do you want to do 2.? (happy to guide you here a bit if you have questions. Also we need to discuss the design here a bit offline maybe)

3. & 5. @pcuenca do you want to take this? (think 3. is more important here)

The other parts we can see tomorrow maybe :-)

[patil-suraj]

ModelMixin should be made state-less by default. E.g. weights will not be saved. Also contrary to transformers should we maybe only work with flax.linen.Module classes here @patil-suraj - I don't really think we need the UNetConditionModel and UNetConditionModule design here - we could just go for class UNetConditionModel(nn.Module): here and make sure everything stays stateless no?

Yes, let's only work with flax.linen.Module, adding the wrapper indeed creates some issues when using these models as submodules in other models.

[patil-suraj]

For 4. we could adapt this https://github.com/patil-suraj/stable-diffusion-jax/blob/main/stable_diffusion_jax/convert_diffusers_to_jax.py#L94

[mishig25]

Yes, I'd like to take point 2 and happy to discuss the details 👍

[pcuenca]

We also need to create a Flax notebook after all is done. I wrote a draft that uses @patil-suraj's repo.

[pcuenca]

Sure, I'll get started with 3.

[kashif]

I am gonna do 5

[patrickvonplaten]

Asked Flax team about design here: google/flax#2454

[pcuenca]

I did some changes in @patil-suraj's repo to easier jit the safety checker, they could be useful when we migrate it to diffusers: patil-suraj/stable-diffusion-jax#11

[dhruvrnaik]

@patrickvonplaten are there any open items for Jax/Flax integration? I'd love to contribute

[pcuenca]

I think this should be complete now, please @patil-suraj @patrickvonplaten feel free to reopen otherwise.