import os
import gc
import torch
import torch._dynamo
from typing import TypeAlias
from torch import Generator
from PIL.Image import Image
from diffusers import (
    FluxPipeline,
    FluxTransformer2DModel,
    AutoencoderTiny,
    DiffusionPipeline,
)
from huggingface_hub.constants import HF_HUB_CACHE
from pipelines.models import TextToImageRequest
from torchao.quantization import quantize_, int8_weight_only
from transformers import T5EncoderModel

torch._dynamo.config.suppress_errors = True

Pipeline: TypeAlias = FluxPipeline

torch.backends.cudnn.benchmark = True
torch._inductor.config.conv_1x1_as_mm = True
torch._inductor.config.coordinate_descent_tuning = True
torch._inductor.config.epilogue_fusion = False
torch._inductor.config.coordinate_descent_check_all_directions = True

os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
os.environ["TOKENIZERS_PARALLELISM"] = "True"

CHECKPOINT = "winner632/flux1-schnell-int8wo"
REVISION = "d9ff2fc9ad81476d3ef3a5f40d273f0fa5a36f2b"


def clear_gpu_cache():
    """Frees GPU memory to prevent memory leaks."""
    gc.collect()
    torch.cuda.empty_cache()
    torch.cuda.reset_max_memory_allocated()
    torch.cuda.reset_peak_memory_stats()


def load_pipeline() -> Pipeline:
    """Loads the diffusion pipeline with quantization and optimizations."""
    clear_gpu_cache()

    transformer_model = FluxTransformer2DModel.from_pretrained(
        os.path.join(
            HF_HUB_CACHE,
            "models--winner632--flux1-schnell-int8wo/snapshots/d9ff2fc9ad81476d3ef3a5f40d273f0fa5a36f2b/transformer",
        ),
        use_safetensors=True,
        local_files_only=True,
        torch_dtype=torch.bfloat16,
    )


    pipe = FluxPipeline.from_pretrained(
        CHECKPOINT,
        revision=REVISION,
        transformer=transformer_model,
        local_files_only=True,
        torch_dtype=torch.bfloat16,
    ).to("cuda")

    pipe.to(memory_format=torch.channels_last)
    pipe.transformer = torch.compile(pipe.transformer, mode="reduce-overhead")

    quantize_(pipe.transformer, int8_weight_only())
    quantize_(pipe.vae, int8_weight_only())
    pipe.vae = torch.compile(pipe.vae, mode="reduce-overhead")

    with torch.no_grad():
        for _ in range(5):
            pipe(
                prompt="onomancy, aftergo, spirantic, Platyhelmia, modificator, drupaceous, jobbernowl, hereness",
                width=1024,
                height=1024,
                guidance_scale=0,
                num_inference_steps=4,
                max_sequence_length=256,
            )

    clear_gpu_cache()
    return pipe


@torch.no_grad()
def infer(
    request: TextToImageRequest, pipeline: Pipeline, generator: Generator
) -> Image:
    """Generates an image from text input using the loaded pipeline."""
    return pipeline(
        request.prompt,
        generator=generator,
        guidance_scale=0e0,
        num_inference_steps=4,
        max_sequence_length=256,
        height=request.height,
        width=request.width,
        output_type="pil",
    ).images[0]


# Example Usage
if __name__ == "__main__":
    print("load pipeline...")
    diffusion_pipeline = load_pipeline()

    sample_request = TextToImageRequest(
        prompt="A futuristic cityscape with neon lights",
        height=1024,
        width=1024,
    )

    generator = torch.Generator(device="cuda").manual_seed(42)

    print("Generating image...")
    generated_img = infer(sample_request, diffusion_pipeline, generator)
    generated_img.show()