import gradio as gr
import torch
from transformers import AutoModelForCausalLM, GemmaTokenizerFast, StoppingCriteria, StoppingCriteriaList, TextIteratorStreamer
from threading import Thread

# Load tokenizer and model
tokenizer = GemmaTokenizerFast.from_pretrained("buddhist-nlp/gemma2-mitra-bo-instruct")
model = AutoModelForCausalLM.from_pretrained("buddhist-nlp/gemma2-mitra-bo-instruct", torch_dtype=torch.float16).to('cuda:0')

# Define custom stopping criteria
class StopOnTokens(StoppingCriteria):
    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
        # Define stop tokens (adjust based on your model's tokenizer)
        stop_ids = [29, 0]  # These should be the token IDs for end of response or similar tokens
        for stop_id in stop_ids:
            if input_ids[0][-1] == stop_id:
                return True
        return False

# Define prediction function for the chat interface
def predict(message, history):
    # Format the input according to your specified structure
    formatted_input = f"### user : {message}  ### input: ### answer:"

    # Tokenize the input
    model_inputs = tokenizer([formatted_input], return_tensors="pt").to("cuda")
    
    # Set up the streamer for partial message output
    streamer = TextIteratorStreamer(tokenizer, timeout=10., skip_prompt=True, skip_special_tokens=True)
    
    # Generate settings
    generate_kwargs = dict(
        model_inputs,
        streamer=streamer,
        max_new_tokens=1024
    )
    
    # Run generation in a separate thread
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()

    # Stream partial messages as they are generated
    partial_message = ""
    for new_token in streamer:
        if new_token != '<':  # Skip specific tokens if necessary
            partial_message += new_token
            yield partial_message

# Create the chat interface using Gradio
gr.ChatInterface(fn=predict, title="Monlam LLM", description="").launch(share=True)