app.py

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

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

model_id = "thrishala/mental_health_chatbot"

try:
    # Load model with appropriate settings
    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        device_map="auto",
        torch_dtype=torch.float16,
        low_cpu_mem_usage=True,
        max_memory={0: "15GiB"} if torch.cuda.is_available() else None,
        offload_folder="offload",
    ).eval()

    tokenizer = AutoTokenizer.from_pretrained(model_id)
    tokenizer.pad_token = tokenizer.eos_token
    tokenizer.model_max_length = 4096  # Set to model's actual context length

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

def generate_text_streaming(prompt, max_new_tokens=128):
    inputs = tokenizer(
        prompt, 
        return_tensors="pt", 
        truncation=True, 
        max_length=4096  # Match model's context length
    ).to(model.device)
    
    generated_tokens = []
    with torch.no_grad():
        for _ in range(max_new_tokens):
            outputs = model.generate(
                **inputs,
                max_new_tokens=1,
                do_sample=False,
                eos_token_id=tokenizer.eos_token_id,
                return_dict_in_generate=True
            )
            
            new_token = outputs.sequences[0, -1]
            generated_tokens.append(new_token)
            
            # Update inputs for next iteration
            inputs = {
                "input_ids": torch.cat([inputs["input_ids"], new_token.unsqueeze(0).unsqueeze(0)], dim=-1),
                "attention_mask": torch.cat([inputs["attention_mask"], torch.ones(1, 1, device=model.device)], dim=-1)
            }
            
            # Decode the accumulated tokens
            current_text = tokenizer.decode(generated_tokens, skip_special_tokens=True)
            yield current_text  # Yield the full text so far
            
            if new_token == tokenizer.eos_token_id:
                break

def respond(message, history, system_message, max_tokens):
    # Build prompt with full history
    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:"
    
    # Keep track of the full response
    full_response = ""
    
    try:
        for token_chunk in generate_text_streaming(prompt, max_tokens):
            # Update the full response and yield incremental changes
            full_response = token_chunk
            yield full_response
            
    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=512, value=128, step=1, label="Max new tokens"),
    ],
)

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