src/pipeline.py

import os
import gc
import torch
import numpy as np
from PIL import Image
from typing import Optional

from diffusers import (
    DiffusionPipeline,
    AutoencoderKL,
    FluxPipeline,
    FluxTransformer2DModel
)
from diffusers.image_processor import VaeImageProcessor
from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
from huggingface_hub.constants import HF_HUB_CACHE
from transformers import (
    T5EncoderModel,
    T5TokenizerFast,
    CLIPTokenizer,
    CLIPTextModel
)
from pipelines.models import TextToImageRequest
from torch import Generator
from torchao.quantization import quantize_, int8_weight_only

# Pre-configurations
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

# Global variables
Pipeline = None
CKPT_ID = "black-forest-labs/FLUX.1-schnell"
CKPT_REVISION = "741f7c3ce8b383c54771c7003378a50191e9efe9"


def empty_cache():
    """Utility function to clear GPU memory."""
    gc.collect()
    torch.cuda.empty_cache()
    torch.cuda.reset_max_memory_allocated()
    torch.cuda.reset_peak_memory_stats()


def load_pipeline() -> FluxPipeline:
    """Loads the diffusion pipeline with specified models and configurations."""
    # Load text encoder
    text_encoder_2 = T5EncoderModel.from_pretrained(
        "Chrissy1/extra0manQ0",
        revision="c0db1e82d89825a4664ad873f20d261cbe46e737",
        subfolder="text_encoder_2",
        torch_dtype=torch.bfloat16
    ).to(memory_format=torch.channels_last)

    # Load transformer
    transformer_path = os.path.join(
        HF_HUB_CACHE,
        "models--Chrissy1--extra0manQ0/snapshots/c0db1e82d89825a4664ad873f20d261cbe46e737/transformer"
    )
    transformer = FluxTransformer2DModel.from_pretrained(
        transformer_path, 
        torch_dtype=torch.bfloat16, 
        use_safetensors=False
    ).to(memory_format=torch.channels_last)

    # Load and quantize autoencoder
    vae = AutoencoderKL.from_pretrained(
        CKPT_ID,
        revision=CKPT_REVISION,
        subfolder="vae",
        local_files_only=True,
        torch_dtype=torch.bfloat16
    )
    quantize_(vae, int8_weight_only())

    # Load FluxPipeline
    pipeline = FluxPipeline.from_pretrained(
        CKPT_ID,
        revision=CKPT_REVISION,
        transformer=transformer,
        text_encoder_2=text_encoder_2,
        torch_dtype=torch.bfloat16
    )
    pipeline.to("cuda")

    # Warm-up run to ensure the pipeline is ready
    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: FluxPipeline, generator: Generator) -> Image:
    """Generates an image based on the input request and pipeline."""
    empty_cache()  # Clear cache before inference
    
    result = 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"
    )
    
    return result.images[0]