src/pipeline.py

from huggingface_hub.constants import HF_HUB_CACHE
from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel
import torch
import torch._dynamo
import gc
import os
from diffusers import FluxPipeline, AutoencoderKL, AutoencoderTiny
from PIL.Image import Image
from pipelines.models import TextToImageRequest
from torch import Generator
from diffusers import FluxTransformer2DModel, DiffusionPipeline
from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only

Pipeline = None

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

torch._dynamo.config.suppress_errors = True

def load_pipeline() -> Pipeline:
    quantize_(AutoencoderKL.from_pretrained("black-forest-labs/FLUX.1-schnell",revision="741f7c3ce8b383c54771c7003378a50191e9efe9", subfolder="vae", torch_dtype=torch.bfloat16, local_files_only=True), int8_weight_only())

    transformer = FluxTransformer2DModel.from_pretrained(os.path.join(HF_HUB_CACHE, "models--edgetensor--edgetensor-FLUX.1-schnell-int8wo/snapshots/6b4c594d4c510da13a40f8c3b483789eb82d36df"), torch_dtype=torch.bfloat16, use_safetensors=False).to(memory_format=torch.channels_last)
    text_encoder_2 = T5EncoderModel.from_pretrained("edgetensor/edgetensor-t5-v1_1-xxl-encoder-bf16", revision = "bfed5213335c2ead9c9b5aff657680db420a7c7d", torch_dtype=torch.bfloat16).to(memory_format=torch.channels_last)
    
    pipeline = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", revision="741f7c3ce8b383c54771c7003378a50191e9efe9", transformer=transformer, torch_dtype=torch.bfloat16, text_encoder_2=text_encoder_2)
    pipeline.to("cuda")

    for idx in range(2):
        pipeline(prompt="satiety, acuity, unwitherable, Pygmy, ramlike, Curtis, fingerstone, rewhisper", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
    return pipeline

@torch.no_grad()
def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
    generator = Generator(pipeline.device).manual_seed(request.seed)

    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]