src/pipeline.py

from huggingface_hub.constants import HF_HUB_CACHE
from pipelines.models import TextToImageRequest
from torch import Generator
from diffusers import FluxTransformer2DModel, DiffusionPipeline
from torchao.quantization import quantize_, int8_weight_only
from transformers import T5EncoderModel
from diffusers import AutoencoderKL
from PIL.Image import Image
import torch
import torch._dynamo
import os

os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
os.environ["TOKENIZERS_PARALLELISM"] = "True"
torch._dynamo.config.suppress_errors = True

Pipeline = None
ids = "black-forest-labs/FLUX.1-schnell"
Revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"

def load_pipeline() -> Pipeline:
    quantize_(AutoencoderKL.from_pretrained(ids,revision=Revision, subfolder="vae", local_files_only=True, torch_dtype=torch.bfloat16,), int8_weight_only())
    
    text_encoder_2 = T5EncoderModel.from_pretrained("agentbot/t5-v1_1-xxl-encoder-bf16_", revision = "208e3686b3027985dbd8c9098c273e0155c77ef4", torch_dtype=torch.bfloat16).to(memory_format=torch.channels_last)
    transformer = FluxTransformer2DModel.from_pretrained(os.path.join(HF_HUB_CACHE, "models--agentbot--FLUX.1-schnell-int8wo_/snapshots/aa66177be06aba5a88dbe7265255bec48833a936"), torch_dtype=torch.bfloat16, use_safetensors=False).to(memory_format=torch.channels_last)
    
    pipeline = DiffusionPipeline.from_pretrained(ids, revision=Revision, transformer=transformer, text_encoder_2=text_encoder_2, torch_dtype=torch.bfloat16,)
    pipeline.to("cuda")
    pipeline(prompt="satiety, 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:
    return pipeline(
        request.prompt,
        generator=Generator(pipeline.device).manual_seed(request.seed),
        guidance_scale=0.0,
        num_inference_steps=4,
        max_sequence_length=256,
        height=request.height,
        width=request.width,
    ).images[0]