Add ONNX conversion pass support for diffusers model

[xiaoyu-work]

Describe your changes

Add ONNX conversion pass support for diffusers model. The exported model will be a CompositeModel.

The model can be exported with CLI:

olive capture-onnx-graph -m "runwayml/stable-diffusion-v1-5" -o "sd15"

or config file with OnnxConversion pass.

The exported model can be loaded with from optimum.onnxruntime import ORTStableDiffusionPipeline directly.

import onnxruntime as ort
import torch
from pathlib import Path
from optimum.onnxruntime import ORTStableDiffusionPipeline
from diffusers import DiffusionPipeline, OnnxRuntimeModel

ort.set_default_logger_severity(3)

sess_options = ort.SessionOptions()
sess_options.enable_mem_pattern = False

model_id = "runwayml/stable-diffusion-v1-5"
onnx_model_path = Path("sd15")
provider = "CUDAExecutionProvider"
pipeline = DiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float32)

vae_encoder_session = OnnxRuntimeModel.load_model(onnx_model_path / "vae_encoder" / "model.onnx", provider=provider)
vae_decoder_session = OnnxRuntimeModel.load_model(onnx_model_path / "vae_decoder" / "model.onnx", provider=provider)
text_encoder_session = OnnxRuntimeModel.load_model(onnx_model_path / "text_encoder" / "model.onnx", provider=provider)
unet_session = OnnxRuntimeModel.load_model(onnx_model_path / "unet" / "model.onnx", provider=provider)

onnx_pipeline = ORTStableDiffusionPipeline(
    vae_encoder_session=vae_encoder_session,
    vae_decoder_session=vae_decoder_session,
    text_encoder_session=text_encoder_session,
    unet_session=unet_session,
    tokenizer=pipeline.tokenizer,
    scheduler=pipeline.scheduler,
    feature_extractor=pipeline.feature_extractor,
    config=dict(pipeline.config),
)

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all tests can pass.
• Update documents if necessary.
• Lint and apply fixes to your code by running lintrunner -a
• Is this a user-facing change? If yes, give a description of this change to be included in the release notes.

(Optional) Issue link