from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel, PeftConfig
import gradio as gr
from huggingface_hub import login
import torch
import os

hf_token = os.getenv("llama")
print(hf_token)

assert hf_token is not None, "Token is missing! Make sure 'llama' is set in the environment."
#login(hf_token)
# Model and adapter paths
model_name = "unsloth/llama-3.2-1b-instruct-bnb-4bit"  # Base model
adapter_name = "Alkhalaf/lora_model"  # LoRA model adapter

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name, token=hf_token)

# Load the LoRA adapter configuration
peft_config = PeftConfig.from_pretrained(adapter_name, token=hf_token)

# Load the base model
base_model = AutoModelForCausalLM.from_pretrained(
    peft_config.base_model_name_or_path,
    token=hf_token,
   
    #torch_dtype=torch.float16

    
)
# Apply the LoRA adapter to the base model
model = PeftModel.from_pretrained(base_model, adapter_name, token=hf_token)

# Define prediction function
def predict(input_text):
    inputs = tokenizer(input_text, return_tensors="pt")
    outputs = model.generate(inputs["input_ids"], max_length=150)
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return response

# Gradio Interface
interface = gr.Interface(
    fn=predict,
    inputs="text",
    outputs="text",
    title="Conversational AI with LoRA",
    description="Interact with a fine-tuned LoRA model for conversational AI."
)

if __name__ == "__main__":
    interface.launch(share=True)