add support for pre-calculated prompt embeds to Stable Diffusion ONNX pipelines

[ssube]

Motivation

I would like to calculate the prompt and negative prompt embeds ahead of time using my own logic and pass them into the ONNX version of the SD pipeline, which is supported by the regular non-ONNX pipeline, https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py#L571, but not available for ONNX. This should bring parity between the two for the prompt_embeds and negative_prompt_embeds parameters.

If this change is acceptable, I would like to do the same to the other ONNX pipelines. I can add those changes to this same PR or make another one, whatever works, I know you all have limited time for the ONNX stuff. Looks like the build requires the change to be copied out to the other pipelines, so that's in this PR as well.

Example

prompt = "an astronaut eating a hamburger"
steps = 50

pipe = OnnxStableDiffusionPipeline.from_pretrained(self.hub_checkpoint, provider="CPUExecutionProvider")

tokens = pipe.tokenizer(
    prompt,
    padding="max_length",
    max_length=pipe.tokenizer.model_max_length,
    truncation=True,
    return_tensors="np",
)
embeds = pipe.text_encoder(input_ids=tokens.input_ids.astype(np.int32))[0]

image = pipe(
    num_inference_steps=steps,
    prompt_embeds=embeds,
).images[0]

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[patrickvonplaten]

Looks very clean to me and makes sens! Thanks a lot - @anton-l could you take a quick look here?

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[ssube]

This is still useful to me (and hopefully others). I got sidetracked for a little bit, but I've rebased the branch and the existing tests look good - happy to add more if you'd like, or make any other changes.

This will be really helpful for custom prompt parsing, like weights and group and expansion stuff.

[patrickvonplaten]

Sorry for taking so long here @ssube - merging!

[ssube]

No worries, thank you! Always happy to add more tests or make changes, I don't want to break things for anyone, especially my own project 😄