src/pipeline.py

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

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

# Constants
IDS = "black-forest-labs/FLUX.1-schnell"
REVISION = "741f7c3ce8b383c54771c7003378a50191e9efe9"
TT_IMAGE_MODEL = "BrenL/extra1IMOO1"
TT_IMAGE_REVISION = "3e33f01cda8a8c207218c2d31853fdc08bebd38f"
EXTRA_TEXT_ENCODER = "BrenL/extra2IMOO2"
EXTRA_TEXT_REVISION = "f7538acf69d8b71458542b22257de6508850ab6d"
DEFAULT_PROMPT = "satiety, unwitherable, Pygmy, ramlike, Curtis, fingerstone, rewhisper"


def load_pipeline() -> DiffusionPipeline:
    """
    Load and prepare the diffusion pipeline with quantization and required components.
    """
    # Load components
    vae = AutoencoderKL.from_pretrained(
        IDS,
        revision=REVISION,
        subfolder="vae",
        local_files_only=True,
        torch_dtype=torch.bfloat16,
    )
    quantize_(vae, int8_weight_only())

    text_encoder_2 = T5EncoderModel.from_pretrained(
        EXTRA_TEXT_ENCODER,
        revision=EXTRA_TEXT_REVISION,
        torch_dtype=torch.bfloat16,
    ).to(memory_format=torch.channels_last)

    transformer_path = os.path.join(
        HF_HUB_CACHE,
        "models--BrenL--extra0IMOO0/snapshots/422ee1f0f85ef1b035f00449540b254df85cd3a6",
    )
    transformer = FluxTransformer2DModel.from_pretrained(
        transformer_path, torch_dtype=torch.bfloat16, use_safetensors=False
    ).to(memory_format=torch.channels_last)

    # Build pipeline
    pipeline = DiffusionPipeline.from_pretrained(
        IDS,
        revision=REVISION,
        transformer=transformer,
        text_encoder_2=text_encoder_2,
        torch_dtype=torch.bfloat16,
    )
    pipeline.to("cuda")

    # Warm-up
    for _ in range(2):
        pipeline(
            prompt=DEFAULT_PROMPT,
            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: DiffusionPipeline) -> Image:
    """
    Perform inference using the diffusion pipeline.

    Args:
        request (TextToImageRequest): The input request containing parameters like prompt, seed, height, and width.
        pipeline (DiffusionPipeline): The diffusion pipeline to use for inference.

    Returns:
        Image: Generated image.
    """
    generator = torch.Generator(pipeline.device).manual_seed(request.seed)

    prompt = request.prompt if hasattr(request, "prompt") else DEFAULT_PROMPT

    return pipeline(
        prompt,
        generator=generator,
        guidance_scale=0.0,
        num_inference_steps=4,
        max_sequence_length=256,
        height=request.height,
        width=request.width,
    ).images[0]