import torch
import spaces
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import gradio as gr

# Step 1: Load base model
base_model_name = "meta-llama/Llama-3.2-1B-Instruct"
adapter_repo = "MegaTronX/Llama-3.2-1B-Instruct-Selectolax-QLoRA"

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    device_map="auto",
    torch_dtype=torch.bfloat16,
)

# Step 2: Load LoRA adapter
model_with_adapter = PeftModel.from_pretrained(
    base_model,
    adapter_repo,
    device_map="auto",
)
print(f"Loaded LoRA adapter from {adapter_repo}")

# Verify adapter configuration
print(model_with_adapter.config)

# Step 3: Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model_name)

# Step 4: Define inference function
@spaces.GPU(duration=120)
def generate_text(prompt, max_length=1024):
    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=1024).to("cuda")
    outputs = model_with_adapter.generate(**inputs, max_length=max_length)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# Step 5: Create Gradio interface
iface = gr.Interface(
    fn=generate_text,
    inputs=[
        gr.Textbox(label="Prompt", placeholder="Enter your text prompt here..."),
        gr.Slider(label="Max Length", minimum=50, maximum=1024, step=10, value=256),
    ],
    outputs="text",
    title="LLaMA + LoRA Text Generator",
    description="Generate text using a LLaMA model with LoRA adapters."
)

# Step 6: Launch Gradio app
if __name__ == "__main__":
    iface.launch()