app.py

import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

from config import load_config

config = load_config("config.yaml")

model_config = config["model_config"]

model_name = model_config.pop("model_name")

# Convert torch_dtype from string to torch.dtype
if "torch_dtype" in model_config:
    if model_config["torch_dtype"] == "float32":
        model_config["torch_dtype"] = torch.float32
    elif model_config["torch_dtype"] == "float16":
        model_config["torch_dtype"] = torch.float16
    elif model_config["torch_dtype"] == "bfloat16":
        model_config["torch_dtype"] = torch.bfloat16

model = AutoModelForCausalLM.from_pretrained(model_name, **model_config)

checkpoint_model = "checkpoint_dir/checkpoint-650"

model.load_adapter(checkpoint_model)


tokenizer = AutoTokenizer.from_pretrained(checkpoint_model, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)


def respond(
    message,
    history,
    system_message,
    max_tokens=256,
    temperature=0.3,
    top_p=0.95,
):
    # Construct the chat list
    chat_list = [{"role": "system", "content": system_message}]
    for user, assistant in history:
        chat_list.extend(
            [
                {"role": "user", "content": user},
                {"role": "assistant", "content": assistant},
            ]
        )
    chat_list.append({"role": "user", "content": message})

    # Manually construct the prompt
    prompt = ""
    for chat in chat_list:
        prompt += f"{chat['role']}: {chat['content']}\n"
    prompt += "assistant:"

    # Get the input length
    input_ids = pipe.tokenizer.encode(prompt, return_tensors="pt")
    input_length = input_ids.shape[1]

    # Adjust max_length to be greater than input_length
    max_length = input_length + max_tokens

    # Ensure max_length is an integer
    max_length = int(max_length)

    outputs = pipe(
        prompt,
        max_new_tokens=int(max_tokens),  # Ensure this is an integer
        max_length=max_length,
        num_beams=1,
        do_sample=True,
        temperature=temperature,
        top_p=top_p,
        top_k=50
    )
    new_text = outputs[0]['generated_text'][len(prompt):]
    return new_text.strip()



examples = [
    ["Suggest some breeds that get along with each other"],
    ["Explain LLM in AI"],
    ["I want to explore Dubai. What are the best places to visit?"],
]

demo = gr.ChatInterface(
    respond,
    textbox=gr.Textbox(
        placeholder="Enter your message here...", container=False, scale=7
    ),
    examples=examples,
    additional_inputs=[
        gr.Textbox(
            value="You are General Knowledge Assistant. Answer the questions based on the provided information. Be succinct and use first-principles thinking to answer the questions.",
            label="System message",
        ),
        gr.Slider(minimum=1, maximum=2048, value=256, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=2.0, value=0.3, step=0.1, label="Temperature"),
        gr.Slider(
            minimum=0.1,
            maximum=1.0,
            value=0.95,
            step=0.05,
            label="Top-p (nucleus sampling)",
        ),
    ],
    title="General Knowledge Assistant",
    description="Ask me anything about general knowledge. I'll try to answer succinctly using first principles.",
)


if __name__ == "__main__":
    demo.launch(debug=True)