app.py

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

if torch.cuda.is_available():
    device = "cuda"
else:
    device = "cpu"

model_id = "thrishala/mental_health_chatbot"

try:
    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        device_map=device,  # Use the determined device
        torch_dtype=torch.float16,
        low_cpu_mem_usage=True,
        max_memory={device: "15GB"},  # Use device-specific memory management
        offload_folder="offload",
    )
    tokenizer = AutoTokenizer.from_pretrained(model_id)
    tokenizer.model_max_length = 512  # Set maximum length

except Exception as e:
    print(f"Error loading model: {e}")
    exit()

def generate_text(prompt, max_new_tokens=128):
    input_ids = tokenizer.encode(prompt, return_tensors="pt").to(model.device)

    with torch.no_grad():
        output = model.generate(
            input_ids=input_ids,
            max_new_tokens=max_new_tokens,
            do_sample=False,  # Or True for sampling
            eos_token_id=tokenizer.eos_token_id,
        )

    generated_text = tokenizer.decode(output[0], skip_special_tokens=True)
    return generated_text

def generate_text_streaming(prompt, max_new_tokens=128):
    input_ids = tokenizer.encode(prompt, return_tensors="pt").to(model.device)

    with torch.no_grad():
        for i in range(max_new_tokens):
            output = model.generate(
                input_ids=input_ids,
                max_new_tokens=1,  # Generate only 1 new token at a time
                do_sample=False,  # Or True for sampling
                eos_token_id=tokenizer.eos_token_id,
                return_dict=True, #Return a dictionary
                output_scores=True #Return the scores
            )

            generated_token = tokenizer.decode(output.logits[0][-1].argmax(), skip_special_tokens=True) #Decode the last token only
            yield generated_token #Yield the last token

            input_ids = torch.cat([input_ids, output.sequences[:, -1:]], dim=-1) #Append the new token to the input

            if output.sequences[0][-1] == tokenizer.eos_token_id: #Check if the end of sequence token was generated
                break #Break the loop

def respond(message, history, system_message, max_tokens):
    prompt = f"{system_message}\n"
    for user_msg, bot_msg in history:
        prompt += f"User: {user_msg}\nAssistant: {bot_msg}\n"
    prompt += f"User: {message}\nAssistant:"

    try:
        for token in generate_text_streaming(prompt, max_tokens): #Iterate over the generator
            yield token #Yield each token individually

    except Exception as e:
        print(f"Error during generation: {e}")
        yield "An error occurred."

demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Textbox(
            value="You are a friendly and helpful mental health chatbot.",
            label="System message",
        ),
        gr.Slider(minimum=1, maximum=128, value=32, step=10, label="Max new tokens"),
    ],
)

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