src/pipeline.py

import torch
#import xformers
#import triton
from PIL.Image import Image
from onediffx.deep_cache import StableDiffusionXLPipeline
#from diffusers import StableDiffusionXLPipeline
from pipelines.models import TextToImageRequest
from torch import Generator
from diffusers import DDIMScheduler
from loss import SchedulerWrapper

#from onediff.schedulers import EulerDiscreteScheduler
from onediffx import compile_pipe


def load_pipeline(pipeline=None) -> StableDiffusionXLPipeline:
    if not pipeline:
        pipeline = StableDiffusionXLPipeline.from_pretrained(
            "stablediffusionapi/newdream-sdxl-20",
            torch_dtype=torch.float16,
        )
    pipeline.to("cuda")
    pipeline.scheduler = SchedulerWrapper(DDIMScheduler.from_config(pipeline.scheduler.config))    
    pipeline = compile_pipe(pipeline)
    pipeline.scheduler.prepare_loss()
    for _ in range(4):
        deepcache_output = pipeline(prompt="kamala harris defends my submission", output_type="pil", cache_interval=1, cache_layer_id=1, cache_block_id=0)
    return pipeline

def infer(request: TextToImageRequest, pipeline: StableDiffusionXLPipeline) -> Image:
    if request.seed is None:
        generator = None
    else:
        generator = Generator(pipeline.device).manual_seed(request.seed)

    return pipeline(
        prompt=request.prompt,
        negative_prompt=request.negative_prompt,
        width=request.width,
        height=request.height,
        generator=generator,
        end_cfg=0.5,
        num_inference_steps=14,
        cache_interval=1,
        cache_layer_id=1,
        cache_block_id=0,
        eta=1,
        guidance_scale = 5.0,
        guidance_rescale = 0.0,
    ).images[0]