src/pipeline.py

from diffusers import AutoencoderKL
from diffusers.image_processor import VaeImageProcessor
import torch
import torch._dynamo
import gc
from PIL.Image import Image
from pipelines.models import TextToImageRequest
from torch import Generator
from diffusers import DiffusionPipeline
from torchao.quantization import quantize_, int8_weight_only

Pipeline = None
MODEL_ID = "black-forest-labs/FLUX.1-schnell"
def clear():
    gc.collect()
    torch.cuda.empty_cache()
    torch.cuda.reset_max_memory_allocated()
    torch.cuda.reset_peak_memory_stats()

def load_pipeline() -> Pipeline:    
    clear()
    dtype, device = torch.bfloat16, "cuda"
    vae = AutoencoderKL.from_pretrained(
        MODEL_ID, subfolder="vae", torch_dtype=torch.bfloat16
    )
    quantize_(vae, int8_weight_only(), device="cuda")
    pipeline = DiffusionPipeline.from_pretrained(
        MODEL_ID,
        vae=vae,
        torch_dtype=dtype,
        )
    pipeline.enable_sequential_cpu_offload()
    for _ in range(2):
        pipeline(prompt="unpervaded, unencumber, froggish, groundneedle, transnatural, fatherhood, outjump, cinerator", width=1024, height=1024, guidance_scale=0.1, num_inference_steps=4, max_sequence_length=256)
    clear()
    return pipeline

@torch.inference_mode()
def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
    clear()
    generator = Generator("cuda").manual_seed(request.seed)
    image=pipeline(request.prompt,generator=generator, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256, height=request.height, width=request.width, output_type="pil").images[0]
    return image