app.py

import os
os.environ["TRANSFORMERS_NO_FLASH_ATTN"] = "1"

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

MODEL_NAME = "openbmb/MiniCPM-o-2_6"

print("Loading model...")

model = AutoModel.from_pretrained(
    MODEL_NAME,
    trust_remote_code=True,
    torch_dtype=torch.float32,
    low_cpu_mem_usage=True,
    attn_implementation="eager"   # 🔥 IMPORTANT FIX
)

model = model.eval()

tokenizer = AutoTokenizer.from_pretrained(
    MODEL_NAME,
    trust_remote_code=True
)

# Init TTS safely
try:
    model.init_tts()
    print("TTS ready")
except Exception as e:
    print("TTS not available:", e)

def chat(text):

    msgs = [
        {"role": "user", "content": [text]}
    ]

    audio_path = "output.wav"

    try:
        result = model.chat(
            msgs=msgs,
            tokenizer=tokenizer,
            generate_audio=True,
            output_audio_path=audio_path,
            max_new_tokens=200,
            temperature=0.3
        )

        reply = result if isinstance(result, str) else text

        return reply, audio_path

    except Exception as e:
        return str(e), None


demo = gr.Interface(
    fn=chat,
    inputs=gr.Textbox(label="Type message"),
    outputs=[
        gr.Textbox(label="Reply"),
        gr.Audio(label="Voice Output")
    ],
    title="MiniCPM-o Voice Chatbot (CPU Stable)"
)

demo.launch(server_name="0.0.0.0", server_port=7860)