src/pipeline.py

import os
from typing import TypeAlias
import torch
from PIL.Image import Image
from diffusers import FluxPipeline, FluxTransformer2DModel
from huggingface_hub.constants import HF_HUB_CACHE
from pipelines.models import TextToImageRequest
from torch import Generator
from torchao.quantization import quantize_, int8_weight_only

Pipeline = None

os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
os.environ["TOKENIZERS_PARALLELISM"] = "True"

def load_pipeline() -> Pipeline:
    path = os.path.join(HF_HUB_CACHE, "models--farapart--flow.1-fast/snapshots/59ebc4a11e1a6d4fe2085988028c5252f3a07b74/transformer") 
    transformer = FluxTransformer2DModel.from_pretrained(path, use_safetensors=False, local_files_only=True, torch_dtype=torch.bfloat16)

    pipeline = FluxPipeline.from_pretrained("farapart/flow.1-fast", revision="59ebc4a11e1a6d4fe2085988028c5252f3a07b74", transformer=transformer, local_files_only=True, torch_dtype=torch.bfloat16).to("cuda")
    pipeline.to(memory_format=torch.channels_last)
    
    with torch.inference_mode():  
        for _ in range(4):
            pipeline(prompt="onomancy, aftergo, spirantic, Platyhelmia, modificator, drupaceous, jobbernowl, hereness", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
    torch.cuda.empty_cache()
    return pipeline

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