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
import torch
import torch._dynamo
import gc
from PIL import Image
from pipelines.models import TextToImageRequest
from torch import Generator
import time
import math
import torch.nn.functional as F
from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only

# preconfigs
import os

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

# globals
Pipeline = None
ckpt_id = "soft987/FLUX.1.schnell-quant2"
ckpt_revision = "6d93094cc0c92f72236c6de41bddf789b8b0b38e"


def empty_cache():
    gc.collect()
    torch.cuda.empty_cache()
    torch.cuda.reset_max_memory_allocated()
    torch.cuda.reset_peak_memory_stats()


def load_pipeline() -> Pipeline:
    vae = AutoencoderKL.from_pretrained(
        ckpt_id,
        revision=ckpt_revision,
        subfolder="vae",
        local_files_only=True,
        torch_dtype=torch.bfloat16,
    )
    quantize_(vae, int8_weight_only())
    text_encoder_2 = T5EncoderModel.from_pretrained(
        "soft987/FLUX1.schnell-full",
        revision="a05d320df4f5795fb4eff2f85ec117e870c078cb",
        subfolder="text_encoder_2",
        torch_dtype=torch.bfloat16,
    )
    path = os.path.join(
        HF_HUB_CACHE,
        "models--soft987--FLUX1.schnell-full/snapshots/a05d320df4f5795fb4eff2f85ec117e870c078cb/transformer",
    )
    transformer = FluxTransformer2DModel.from_pretrained(
        path, torch_dtype=torch.bfloat16, use_safetensors=False
    )
    pipeline = FluxPipeline.from_pretrained(
        ckpt_id,
        revision=ckpt_revision,
        transformer=transformer,
        text_encoder_2=text_encoder_2,
        torch_dtype=torch.bfloat16,
    )
    pipeline.to("cuda")
    pipeline.to(memory_format=torch.channels_last)
    for _ in range(1):
        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


sample = 1


@torch.no_grad()
def infer(
    request: TextToImageRequest, pipeline: Pipeline, generator: Generator
) -> Image:
    global sample
    if not sample:
        sample = 1
        empty_cache()
    try:
        img = 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]
        return img
    except Exception as e:
        return None