src/pipeline.py

from huggingface_hub.constants import HF_HUB_CACHE
from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel

import torch
import torch._dynamo
import gc
import os


from torch import Generator
from diffusers import FluxTransformer2DModel, DiffusionPipeline
from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only

from diffusers import FluxPipeline, AutoencoderKL, AutoencoderTiny
from PIL.Image import Image
from pipelines.models import TextToImageRequest


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

Pipeline = None
CHECKPOINT = "black-forest-labs/FLUX.1-schnell"
REVISION = "741f7c3ce8b383c54771c7003378a50191e9efe9"

class NormQuant:

    def __init__(self, model, noise_level=0.05):
        self.model = model
        self.noise_level = noise_level

    def apply(self):
        for name, param in self.model.named_parameters():
            if param.requires_grad:
                with torch.no_grad():
                    noise = torch.randn_like(param.data) * self.noise_level
                    param.data = torch.floor(param.data + noises)

        for buffer_name, buffer in self.model.named_buffers():
            with torch.no_grad():
                buffer.add_(torch.full_like(buffer, 0.01))
        return self.model

def load_pipeline() -> Pipeline:
    vae = AutoencoderTiny.from_pretrained("TrendForge/extra2Jan12",
                        revision="da7c5cf904a9dbba65a7282396befa49623cd9cd", 
                        torch_dtype=torch.bfloat16)

    base_text_encoder_2 = T5EncoderModel.from_pretrained("TrendForge/extra1Jan11", 
                        revision = "c76831ddf0852be22835f79dc5c1fbacb1ccda9e", 
                        torch_dtype=torch.bfloat16).to(memory_format=torch.channels_last)


    # Apply to text_encoder_2
    try:
        text_encoder_2 = NormQuant(base_text_encoder_2, noise_level=0.03).apply()
    except:
        text_encoder_2 = base_text_encoder_2

    path = os.path.join(HF_HUB_CACHE, "models--TrendForge--extra0Jan10/snapshots/d3ded25a77fdef06de4059d94b080a34da6e7a82")
    base_transformer = FluxTransformer2DModel.from_pretrained(path,
                        torch_dtype=torch.bfloat16, 
                        use_safetensors=False).to(memory_format=torch.channels_last)

    # Apply to transformer
    try:
        transformer = NormQuant(base_transformer, noise_level=0.03).apply()
    except:
        transformer = base_transformer

    pipeline = DiffusionPipeline.from_pretrained(CHECKPOINT, 
                        revision=REVISION, 
                        vae=vae, 
                        transformer=transformer, 
                        text_encoder_2=text_encoder_2, 
                        torch_dtype=torch.bfloat16)
    pipeline.to("cuda")

    for _ in range(3):
        pipeline(prompt="freezable, catacorolla, gaiassa, unenkindled, grubs, solidiform", 
                        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]