import spaces
import gradio as gr
from cartesia_pytorch import ReneLMHeadModel
from transformers import AutoTokenizer
#import subprocess
#subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
# Load model and tokenizer
model = ReneLMHeadModel.from_pretrained("cartesia-ai/Rene-v0.1-1.3b-pytorch").half().cuda()
tokenizer = AutoTokenizer.from_pretrained("allenai/OLMo-1B-hf")

# Define the function to generate text
@spaces.GPU(duration=120)
def generate_text(input_text):
    inputs = tokenizer([input_text], return_tensors="pt")
    outputs = model.generate(inputs.input_ids.cuda(), max_length=50, top_k=100, top_p=0.99)
    out_message = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
    return out_message

# Create Gradio interface
interface = gr.Interface(
    fn=generate_text, 
    inputs="text", 
    outputs="text",
    title="ReneLM Text Generator",
    description="Generate text using ReneLMHeadModel from a prompt."
)

# Launch the Gradio app
interface.launch()