src/pipeline.py

import os
import gc
import torch
from PIL.Image import Image
from dataclasses import dataclass
from diffusers import DiffusionPipeline, AutoencoderTiny, FluxTransformer2DModel
from transformers import T5EncoderModel
from huggingface_hub.constants import HF_HUB_CACHE
from torchao.quantization import quantize_, int8_weight_only, float8_weight_only
from first_block_cache.diffusers_adapters import apply_cache_on_pipe
from caching import apply_cache_on_pipe
from pipelines.models import TextToImageRequest
from torch import Generator

# Configuration settings using a dataclass for clarity
@dataclass
class Config:
    CKPT_ID: str = "black-forest-labs/FLUX.1-schnell"
    CKPT_REVISION: str = "741f7c3ce8b383c54771c7003378a50191e9efe9"
    DEVICE: str = "cuda"
    DTYPE = torch.bfloat16
    PYTORCH_CUDA_ALLOC_CONF: str = "expandable_segments:True"

def _initialize_environment():
    """Set up PyTorch and CUDA environment variables for optimal performance."""
    torch.backends.cuda.matmul.allow_tf32 = True
    torch.backends.cudnn.enabled = True
    torch.backends.cudnn.benchmark = True
    os.environ['PYTORCH_CUDA_ALLOC_CONF'] = Config.PYTORCH_CUDA_ALLOC_CONF

def _clear_gpu_memory():
    """Free up GPU memory to prevent memory-related issues."""
    gc.collect()
    torch.cuda.empty_cache()
    torch.cuda.reset_max_memory_allocated()
    torch.cuda.reset_peak_memory_stats()

def _load_text_encoder_model():
    """Load the text encoder model with specified configuration."""
    return T5EncoderModel.from_pretrained(
        "city96/t5-v1_1-xxl-encoder-bf16",
        revision="1b9c856aadb864af93c1dcdc226c2774fa67bc86",
        torch_dtype=Config.DTYPE
    ).to(memory_format=torch.channels_last)

def _load_vae_model():
    """Load the variational autoencoder (VAE) model with specified configuration."""
    return AutoencoderTiny.from_pretrained(
        "RobertML/FLUX.1-schnell-vae_e3m2",
        revision="da0d2cd7815792fb40d084dbd8ed32b63f153d8d",
        torch_dtype=Config.DTYPE
    )

def _load_transformer_model():
    """Load the transformer model from a specific cached path."""
    transformer_path = os.path.join(
        HF_HUB_CACHE,
        "models--RobertML--FLUX.1-schnell-int8wo/snapshots/307e0777d92df966a3c0f99f31a6ee8957a9857a"
    )
    return FluxTransformer2DModel.from_pretrained(
        transformer_path,
        torch_dtype=Config.DTYPE,
        use_safetensors=False
    ).to(memory_format=torch.channels_last)

def _warmup_pipeline(pipeline):
    """Warm up the pipeline by running it with an empty prompt to initialize internal caches."""
    for _ in range(3):
        pipeline(prompt=" ")

def load_pipeline():
    """
    Load and configure the diffusion pipeline for text-to-image generation.

    Returns:
        DiffusionPipeline: The configured pipeline ready for inference.
    """
    _clear_gpu_memory()

    # Load individual components
    text_encoder = _load_text_encoder_model()
    vae = _load_vae_model()
    transformer = _load_transformer_model()

    # Assemble the diffusion pipeline
    pipeline = DiffusionPipeline.from_pretrained(
        Config.CKPT_ID,
        vae=vae,
        revision=Config.CKPT_REVISION,
        transformer=transformer,
        text_encoder_2=text_encoder,
        torch_dtype=Config.DTYPE,
    ).to(Config.DEVICE)

    # Apply optimizations
    apply_cache_on_pipe(pipeline)
    pipeline.to(memory_format=torch.channels_last)
    pipeline.vae = torch.compile(pipeline.vae, mode="max-autotune")
    quantize_(pipeline.vae, int8_weight_only())
    quantize_(pipeline.vae, float8_weight_only())

    # Warm up the pipeline to ensure readiness
    _warmup_pipeline(pipeline)

    return pipeline

@torch.no_grad()
def infer(request: TextToImageRequest, pipeline: DiffusionPipeline, generator: Generator) -> Image:
    """
    Generate an image from a text prompt using the diffusion pipeline.

    Args:
        request (TextToImageRequest): The request containing the prompt and image parameters.
        pipeline (DiffusionPipeline): The pre-loaded diffusion pipeline.
        generator (Generator): The random seed generator for reproducibility.

    Returns:
        Image: The generated image in PIL format.
    """
    image = pipeline(
        prompt=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]
    return image

# Initialize environment settings when the module is imported
_initialize_environment()

# For compatibility with other scripts, alias load_pipeline as load
load = load_pipeline