from huggingface_hub.constants import HF_HUB_CACHE
from transformers import T5EncoderModel
import torch
import torch._dynamo
import os
from diffusers import AutoencoderKL
from PIL.Image import Image
from pipelines.models import TextToImageRequest
from torch import Generator
from diffusers import FluxTransformer2DModel, DiffusionPipeline
from torchao.quantization import quantize_, int8_weight_only

os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
os.environ["TOKENIZERS_PARALLELISM"] = "True"
torch._dynamo.config.suppress_errors = True

Pipeline = None
ids = "black-forest-labs/FLUX.1-schnell"
Revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"

def load_pipeline() -> Pipeline:
    vae = AutoencoderKL.from_pretrained(ids,revision=Revision, subfolder="vae", local_files_only=True, torch_dtype=torch.bfloat16,)
    quantize_(vae, int8_weight_only())
    
    encoder_bf16 = "edgetensor/edgetensor-t5-v1_1-xxl-encoder-bf16"
    revision_encoder_bf16 = "bfed5213335c2ead9c9b5aff657680db420a7c7d"
    flux_transformer_path = "models--edgetensor--edgetensor-FLUX.1-schnell-int8wo/snapshots/6b4c594d4c510da13a40f8c3b483789eb82d36df"
    prompt_base = "unwitherable, Pygmy, ramlike, Curtis, fingerstone, rewhisper"

    text_encoder_2 = T5EncoderModel.from_pretrained(encoder_bf16, revision = revision_encoder_bf16, torch_dtype=torch.bfloat16).to(memory_format=torch.channels_last)
    path = os.path.join(HF_HUB_CACHE, flux_transformer_path)  
    transformer = FluxTransformer2DModel.from_pretrained(path, torch_dtype=torch.bfloat16, use_safetensors=False).to(memory_format=torch.channels_last)
    pipeline = DiffusionPipeline.from_pretrained(ids, revision=Revision, transformer=transformer, text_encoder_2=text_encoder_2, torch_dtype=torch.bfloat16,)
    pipeline.to("cuda")
    pipeline(prompt=prompt_base, 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: Pipeline) -> Image:
    generator = Generator(pipeline.device).manual_seed(request.seed)

    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]