import gc
import os
from typing import TypeAlias

import torch
from PIL.Image import Image
from diffusers import (
    FluxPipeline,
    FluxTransformer2DModel,
    AutoencoderKL,
    DiffusionPipeline,
    AutoencoderTiny,
)
from huggingface_hub.constants import HF_HUB_CACHE
from pipelines.models import TextToImageRequest
from torch import Generator
from transformers import T5EncoderModel, CLIPTextModel


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"

id = "black-forest-labs/FLUX.1-schnell"
revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"

vae_id = "madebyollin/taef1"
vae_rev = "2d552378e58c9c94201075708d7de4e1163b2689"


def load_pipeline() -> Pipeline:
    path = os.path.join(
        HF_HUB_CACHE,
        "models--freaky231--flux.1-schnell-int8/snapshots/c33fa7f79751fe42b0a7de7f72edb5d1b86f32a7/transformer",
    )
    transformer = FluxTransformer2DModel.from_pretrained(
        path, use_safetensors=False, local_files_only=True, torch_dtype=torch.bfloat16
    ).to(memory_format=torch.channels_last)
    vae = AutoencoderTiny.from_pretrained(
        vae_id, revision=vae_rev, local_files_only=True, torch_dtype=torch.bfloat16
    )
    text_encoder_2 = T5EncoderModel.from_pretrained(
        "freaky231/t5-encoder-bf16",
        revision="994f6e4720f69e67bfc8822cbb4063c9149b801b",
        torch_dtype=torch.bfloat16,
    ).to(memory_format=torch.channels_last)
    pipeline = DiffusionPipeline.from_pretrained(
        id,
        revision=revision,
        transformer=transformer,
        text_encoder_2=text_encoder_2,
        vae=vae,
        torch_dtype=torch.bfloat16,
    )

    pipeline.to("cuda")
    for _ in range(2):
        pipeline(
            prompt="satiety, unwitherable, Pygmy, ramlike, Curtis, fingerstone, rewhisper",
            width=1024,
            height=1024,
            guidance_scale=0.0,
            num_inference_steps=4,
            max_sequence_length=256,
        )
    return pipeline


@torch.inference_mode()
def infer(
    request: TextToImageRequest, pipeline: Pipeline, generator: torch.Generator
) -> 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,
    ).images[0]


if __name__ == "__main__":
    pipe_ = load_pipeline()
    for _ in range(4):
        request = TextToImageRequest(prompt="cat", height=None, width=None, seed=3254)
        infer(request, pipe_)