app.py

import gradio as gr
from huggingface_hub import hf_hub_download
import subprocess
import tempfile
import shutil
import os
import spaces

from transformers import T5ForConditionalGeneration, T5Tokenizer
import os

subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)


def install_apex():
    # Install Apex in editable mode from the specified GitHub repository
    cmd = [
        'pip', 'install', '--no-cache-dir', '--no-build-isolation',
        '--config-settings', 'build-option=--cpp_ext', '--config-settings',
        'build-option=--cuda_ext', '-e', 'git+https://github.com/NVIDIA/apex.git'
    ]
    subprocess.run(cmd, check=True)

try:
    import apex
except ModuleNotFoundError:
    print("Apex not found, installing...")
    install_apex()
    # Try to import Apex again after installation
    import apex
    

def download_t5_model(model_id, save_directory):
    # Modelin tokenizer'ını ve modeli indir
    model = T5ForConditionalGeneration.from_pretrained(model_id)
    tokenizer = T5Tokenizer.from_pretrained(model_id)

    # Model ve tokenizer'ı belirtilen dizine kaydet
    if not os.path.exists(save_directory):
        os.makedirs(save_directory)
    model.save_pretrained(save_directory)
    tokenizer.save_pretrained(save_directory)

# Model ID ve kaydedilecek dizin
model_id = "DeepFloyd/t5-v1_1-xxl"
save_directory = "pretrained_models/t5_ckpts/t5-v1_1-xxl"

# Modeli indir
download_t5_model(model_id, save_directory)

def download_model(repo_id, model_name):
    model_path = hf_hub_download(repo_id=repo_id, filename=model_name)
    return model_path

import glob

@spaces.GPU
def run_inference(model_name, prompt_text):
    repo_id = "hpcai-tech/Open-Sora"
    
    # Map model names to their respective configuration files
    config_mapping = {
        "OpenSora-v1-16x256x256.pth": "configs/opensora/inference/16x256x256.py",
        "OpenSora-v1-HQ-16x256x256.pth": "configs/opensora/inference/16x512x512.py",
        "OpenSora-v1-HQ-16x512x512.pth": "configs/opensora/inference/64x512x512.py"
    }
    
    config_path = config_mapping[model_name]
    ckpt_path = download_model(repo_id, model_name)

    # Save prompt_text to a temporary text file
    prompt_file = tempfile.NamedTemporaryFile(delete=False, suffix=".txt", mode='w')
    prompt_file.write(prompt_text)
    prompt_file.close()

    with open(config_path, 'r') as file:
        config_content = file.read()
    config_content = config_content.replace('prompt_path = "./assets/texts/t2v_samples.txt"', f'prompt_path = "{prompt_file.name}"')
    
    with tempfile.NamedTemporaryFile('w', delete=False, suffix='.py') as temp_file:
        temp_file.write(config_content)
        temp_config_path = temp_file.name

    cmd = [
        "torchrun", "--standalone", "--nproc_per_node", "1",
        "scripts/inference.py", temp_config_path,
        "--ckpt-path", ckpt_path
    ]
    subprocess.run(cmd)

    save_dir = "./outputs/samples/"  # Örneğin, inference.py tarafından kullanılan kayıt dizini
    list_of_files = glob.glob(f'{save_dir}/*')
    if list_of_files:
        latest_file = max(list_of_files, key=os.path.getctime)
        return latest_file
    else:
        print("No files found in the output directory.")
        return None

    # Clean up the temporary files
    os.remove(temp_file.name)
    os.remove(prompt_file.name)

def main():
    gr.Interface(
        fn=run_inference,
        inputs=[
            gr.Dropdown(choices=[
                "OpenSora-v1-16x256x256.pth",
                "OpenSora-v1-HQ-16x256x256.pth",
                "OpenSora-v1-HQ-16x512x512.pth"
            ], 
            value="OpenSora-v1-16x256x256.pth",
            label="Model Selection"),
            gr.Textbox(label="Prompt Text", value="Enter prompt text here")
        ],
        outputs=gr.Video(label="Output Video"),
        title="Open-Sora Inference",
        description="Run Open-Sora Inference with Custom Parameters",
    ).launch()

if __name__ == "__main__":
    main()