src/pipeline.py

import os
from typing import Type

import torch
import torch._dynamo
import torch.nn.functional as F
from PIL import Image
from torch import Generator
from diffusers import DiffusionPipeline, FluxTransformer2DModel
from huggingface_hub.constants import HF_HUB_CACHE
from transformers import T5EncoderModel
from pipelines.models import TextToImageRequest

# Configure environment variables
os.environ['PYTORCH_CUDA_ALLOC_CONF'] = "expandable_segments:True"
os.environ["TOKENIZERS_PARALLELISM"] = "True"

torch._dynamo.config.suppress_errors = True
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.enabled = True

Pipeline = None

def load_pipeline() -> Pipeline:
    """
    Load and initialize the Diffusion Pipeline with custom components.

    Returns:
        Pipeline: Initialized diffusion pipeline.
    """
    # Configuration for model checkpoints
    ckpt_id = "black-forest-labs/FLUX.1-schnell"
    ckpt_revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"

    # Load the secondary text encoder
    text_encoder_2 = T5EncoderModel.from_pretrained(
        "VictorTn/extra0izer0",
        revision="ea3cc69c8eba166304100f994e9b6ec9f5179a9d",
        subfolder="text_encoder_2",
        torch_dtype=torch.bfloat16
    )

    # Load the transformer model
    transformer_path = os.path.join(
        HF_HUB_CACHE,
        "models--VictorTn--extra0izer0/snapshots/ea3cc69c8eba166304100f994e9b6ec9f5179a9d/transformer"
    )
    transformer = FluxTransformer2DModel.from_pretrained(
        transformer_path,
        torch_dtype=torch.bfloat16,
        use_safetensors=False
    )

    # Initialize the diffusion pipeline
    pipeline = DiffusionPipeline.from_pretrained(
        ckpt_id,
        revision=ckpt_revision,
        transformer=transformer,
        text_encoder_2=text_encoder_2,
        torch_dtype=torch.bfloat16
    )

    # Move pipeline to GPU and optimize memory format
    pipeline.to("cuda")
    pipeline.to(memory_format=torch.channels_last)

    # Perform a warm-up run
    with torch.inference_mode():
        pipeline(
            prompt="insensible, timbale, pothery, electrovital, actinogram, taxis, intracerebellar, centrodesmus",
            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, generator: Generator) -> Image:
    """
    Perform inference using the provided diffusion pipeline.

    Args:
        request (TextToImageRequest): The text-to-image request containing prompt and image dimensions.
        pipeline (Pipeline): The initialized diffusion pipeline.
        generator (Generator): Random generator for reproducibility.

    Returns:
        Image: Generated PIL 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,
        output_type="pil"
    ).images[0]