app.py

import gradio as gr
import numpy as np
import random

import spaces #[uncomment to use ZeroGPU]
from diffusers import DiffusionPipeline
import torch

from peft import PeftModel, LoraConfig
import os

def get_lora_sd_pipeline(
    ckpt_dir='./lora', 
    base_model_name_or_path=None, 
    dtype=torch.float16, 
    adapter_name="default"
    ):

    unet_sub_dir = os.path.join(ckpt_dir, "unet")
    text_encoder_sub_dir = os.path.join(ckpt_dir, "text_encoder")
    
    if os.path.exists(text_encoder_sub_dir) and base_model_name_or_path is None:
        config = LoraConfig.from_pretrained(text_encoder_sub_dir)
        base_model_name_or_path = config.base_model_name_or_path
    
    if base_model_name_or_path is None:
        raise ValueError("Please specify the base model name or path")
    
    pipe = DiffusionPipeline.from_pretrained(base_model_name_or_path, torch_dtype=dtype)
    before_params = pipe.unet.parameters()
    pipe.unet = PeftModel.from_pretrained(pipe.unet, unet_sub_dir, adapter_name=adapter_name)
    pipe.unet.set_adapter(adapter_name)
    after_params = pipe.unet.parameters()
    print("Parameters changed:", any(torch.any(b != a) for b, a in zip(before_params, after_params)))
    
    if os.path.exists(text_encoder_sub_dir):
        pipe.text_encoder = PeftModel.from_pretrained(pipe.text_encoder, text_encoder_sub_dir, adapter_name=adapter_name)
    
    if dtype in (torch.float16, torch.bfloat16):
        pipe.unet.half()
        pipe.text_encoder.half()
    
    return pipe

def process_prompt(prompt, tokenizer, text_encoder, max_length=77):
    tokens = tokenizer(prompt, truncation=False, return_tensors="pt")["input_ids"]
    chunks = [tokens[:, i:i + max_length] for i in range(0, tokens.shape[1], max_length)]
    
    with torch.no_grad():
        embeds = [text_encoder(chunk.to(text_encoder.device))[0] for chunk in chunks]
    
    return torch.cat(embeds, dim=1)

def align_embeddings(prompt_embeds, negative_prompt_embeds):
    max_length = max(prompt_embeds.shape[1], negative_prompt_embeds.shape[1])
    return torch.nn.functional.pad(prompt_embeds, (0, 0, 0, max_length - prompt_embeds.shape[1])), \
           torch.nn.functional.pad(negative_prompt_embeds, (0, 0, 0, max_length - negative_prompt_embeds.shape[1]))

device = "cuda" if torch.cuda.is_available() else "cpu"

model_id_default = "sd-legacy/stable-diffusion-v1-5"
model_dropdown = ['stabilityai/sdxl-turbo', 'CompVis/stable-diffusion-v1-4', 'sd-legacy/stable-diffusion-v1-5']

model_lora_default = "lora"

if torch.cuda.is_available():
    torch_dtype = torch.float16
else:
    torch_dtype = torch.float32

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1024


# @spaces.GPU #[uncomment to use ZeroGPU]
def infer(
    prompt,
    negative_prompt,
    randomize_seed,
    width=512,
    height=512,
    model_repo_id=model_id_default,
    seed=42,
    guidance_scale=7,
    num_inference_steps=20,
    model_lora_id=model_lora_default,
    lora_scale=0.5,
    progress=gr.Progress(track_tqdm=True),
    ):
        
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)

    generator = torch.Generator().manual_seed(seed)
    
    # добавляем обновление pipe по условию
    if model_repo_id != model_id_default:
        pipe = DiffusionPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype).to(device)
        prompt_embeds = process_prompt(prompt, pipe.tokenizer, pipe.text_encoder)
        negative_prompt_embeds = process_prompt(negative_prompt, pipe.tokenizer, pipe.text_encoder)
        prompt_embeds, negative_prompt_embeds = align_embeddings(prompt_embeds, negative_prompt_embeds)
    else:
        # добавляем lora
        pipe = get_lora_sd_pipeline(ckpt_dir='./' + model_lora_id, base_model_name_or_path=model_id_default, dtype=torch_dtype).to(device)
        prompt_embeds = process_prompt(prompt, pipe.tokenizer, pipe.text_encoder)
        negative_prompt_embeds = process_prompt(negative_prompt, pipe.tokenizer, pipe.text_encoder)
        prompt_embeds, negative_prompt_embeds = align_embeddings(prompt_embeds, negative_prompt_embeds)
        print(f"LoRA adapter loaded: {pipe.unet.active_adapters}")
        print(f"LoRA scale applied: {lora_scale}")
        pipe.fuse_lora(lora_scale=lora_scale)

    # на вызов pipe с эмбеддингами
    params = {
        'prompt_embeds': prompt_embeds,
        'negative_prompt_embeds': negative_prompt_embeds,
        'guidance_scale': guidance_scale,
        'num_inference_steps': num_inference_steps,
        'width': width,
        'height': height,
        'generator': generator,
    }
    
    return pipe(**params).images[0], seed    


examples = [
    "A Elon Mask lady in a Russian embroidered kaftan is sitting on a beautiful carved veranda, holding a cup to her mouth and drinking tea from the cup. With her other hand, the girl holds a saucer. The cup and saucer are painted with gzhel. Next to the girl on the table stands a samovar, and steam can be seen above it.",
    "Elon Mask in a jungle, cold color palette, muted colors, detailed, 8k",
    "An Elon Mask astronaut riding a green horse",
    "A delicious Elon Mask ceviche cheesecake slice",
]

css = """
#col-container {
    margin: 0 auto;
    max-width: 640px;
}
"""

with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown(" # Text-to-Image")

        with gr.Row():
            prompt = gr.Text(
                label="Prompt",
                show_label=False,
                max_lines=1,
                placeholder="Enter your prompt",
                container=False,
            )

            run_button = gr.Button("Run", scale=0, variant="primary")

        result = gr.Image(label="Result", show_label=False)

if __name__ == "__main__":
    demo.launch(ssr=False)