app.py

import gradio as gr
import torch
from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
    SpeechT5Processor,
    SpeechT5ForTextToSpeech,
    SpeechT5HifiGan,
    WhisperProcessor,
    WhisperForConditionalGeneration
)
from datasets import load_dataset
import os
import spaces
import tempfile
import soundfile as sf
import librosa

# --- Configuration ---
HUGGINGFACE_MODEL_ID = "HuggingFaceH4/Qwen2.5-1.5B-Instruct-gkd"
TORCH_DTYPE = torch.bfloat16
MAX_NEW_TOKENS = 512
DO_SAMPLE = True
TEMPERATURE = 0.7
TOP_K = 50
TOP_P = 0.95

TTS_MODEL_ID = "microsoft/speecht5_tts"
TTS_VOCODER_ID = "microsoft/speecht5_hifigan"
STT_MODEL_ID = "openai/whisper-small"

# --- Global Variables ---
tokenizer = None
llm_model = None
tts_processor = None
tts_model = None
tts_vocoder = None
speaker_embeddings = None
whisper_processor = None
whisper_model = None
first_load = True

def load_readme():
    with open("README.md", "r", encoding="utf-8") as f:
        return f.read()

# --- Helper: Split Text Into Chunks ---
def split_text_into_chunks(text, max_chars=400):
    sentences = text.replace("...", ".").split(". ")
    chunks = []
    current_chunk = ""
    for sentence in sentences:
        if len(current_chunk) + len(sentence) + 2 < max_chars:
            current_chunk += ". " + sentence if current_chunk else sentence
        else:
            chunks.append(current_chunk)
            current_chunk = sentence
    if current_chunk:
        chunks.append(current_chunk)
    return [f"{chunk}." for chunk in chunks if chunk.strip()]

# --- Load Models Function ---
@spaces.GPU
def load_models():
    global tokenizer, llm_model, tts_processor, tts_model, tts_vocoder, speaker_embeddings, whisper_processor, whisper_model
    hf_token = os.environ.get("HF_TOKEN")

    # LLM
    if tokenizer is None or llm_model is None:
        try:
            tokenizer = AutoTokenizer.from_pretrained(HUGGINGFACE_MODEL_ID, token=hf_token)
            if tokenizer.pad_token is None:
                tokenizer.pad_token = tokenizer.eos_token
            llm_model = AutoModelForCausalLM.from_pretrained(
                HUGGINGFACE_MODEL_ID,
                torch_dtype=TORCH_DTYPE,
                device_map="auto",
                token=hf_token
            ).eval()
            print("LLM loaded successfully.")
        except Exception as e:
            print(f"Error loading LLM: {e}")

    # TTS
    if tts_processor is None or tts_model is None or tts_vocoder is None:
        try:
            tts_processor = SpeechT5Processor.from_pretrained(TTS_MODEL_ID, token=hf_token)
            tts_model = SpeechT5ForTextToSpeech.from_pretrained(TTS_MODEL_ID, token=hf_token)
            tts_vocoder = SpeechT5HifiGan.from_pretrained(TTS_VOCODER_ID, token=hf_token)
            embeddings = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation", token=hf_token)
            speaker_embeddings = torch.tensor(embeddings[7306]["xvector"]).unsqueeze(0)
            device = llm_model.device if llm_model else 'cpu'
            tts_model.to(device)
            tts_vocoder.to(device)
            speaker_embeddings = speaker_embeddings.to(device)
            print("TTS models loaded.")
        except Exception as e:
            print(f"Error loading TTS: {e}")

    # STT
    if whisper_processor is None or whisper_model is None:
        try:
            whisper_processor = WhisperProcessor.from_pretrained(STT_MODEL_ID, token=hf_token)
            whisper_model = WhisperForConditionalGeneration.from_pretrained(STT_MODEL_ID, token=hf_token)
            device = llm_model.device if llm_model else 'cpu'
            whisper_model.to(device)
            print("Whisper loaded.")
        except Exception as e:
            print(f"Error loading Whisper: {e}")

# --- Generate Response and Audio ---
@spaces.GPU
def generate_response_and_audio(message, history):
    global first_load
    if first_load:
        load_models()
        first_load = False

    global tokenizer, llm_model, tts_processor, tts_model, tts_vocoder, speaker_embeddings

    if tokenizer is None or llm_model is None:
        return [{"role": "assistant", "content": "Error: LLM not loaded."}], None

    messages = history.copy()
    messages.append({"role": "user", "content": message})

    try:
        input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    except:
        input_text = ""
        for item in history:
            input_text += f"{item['role'].capitalize()}: {item['content']}\n"
        input_text += f"User: {message}\nAssistant:"

    try:
        inputs = tokenizer(input_text, return_tensors="pt", padding=True, truncation=True).to(llm_model.device)
        output_ids = llm_model.generate(
            inputs["input_ids"],
            attention_mask=inputs["attention_mask"],
            max_new_tokens=MAX_NEW_TOKENS,
            do_sample=DO_SAMPLE,
            temperature=TEMPERATURE,
            top_k=TOP_K,
            top_p=TOP_P,
            pad_token_id=tokenizer.eos_token_id
        )
        generated_text = tokenizer.decode(output_ids[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True).strip()
    except Exception as e:
        print(f"LLM error: {e}")
        return history + [{"role": "assistant", "content": "I had an issue generating a response."}], None

    audio_path = None
    if None not in [tts_processor, tts_model, tts_vocoder, speaker_embeddings]:
        try:
            device = llm_model.device
            text_chunks = split_text_into_chunks(generated_text)

            full_speech = []
            for chunk in text_chunks:
                tts_inputs = tts_processor(text=chunk, return_tensors="pt", max_length=512, truncation=True).to(device)
                speech = tts_model.generate_speech(tts_inputs["input_ids"], speaker_embeddings, vocoder=tts_vocoder)
                full_speech.append(speech.cpu())

            full_speech_tensor = torch.cat(full_speech, dim=0)

            with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp_file:
                audio_path = tmp_file.name
                sf.write(audio_path, full_speech_tensor.numpy(), samplerate=16000)

        except Exception as e:
            print(f"TTS error: {e}")

    return history + [{"role": "assistant", "content": generated_text}], audio_path

# --- Transcribe Audio ---
@spaces.GPU
def transcribe_audio(filepath):
    global first_load
    if first_load:
        load_models()
        first_load = False

    global whisper_processor, whisper_model
    if whisper_model is None:
        return "Whisper model not loaded."

    try:
        audio, sr = librosa.load(filepath, sr=16000)
        inputs = whisper_processor(audio, sampling_rate=sr, return_tensors="pt").input_features.to(whisper_model.device)
        outputs = whisper_model.generate(inputs)
        return whisper_processor.batch_decode(outputs, skip_special_tokens=True)[0]
    except Exception as e:
        return f"Transcription failed: {e}"

# --- Gradio UI ---
with gr.Blocks() as demo:
    gr.Markdown("# Qwen2.5 Chatbot with Voice Input/Output")
    
    with gr.Tab("Chat"):
        chatbot = gr.Chatbot(type='messages')
        text_input = gr.Textbox(placeholder="Type your message...")
        audio_output = gr.Audio(label="Response Audio", autoplay=True)
        text_input.submit(generate_response_and_audio, [text_input, chatbot], [chatbot, audio_output])

    with gr.Tab("Transcribe"):
        audio_input = gr.Audio(type="filepath", label="Upload Audio")
        transcribed = gr.Textbox(label="Transcription")
        audio_input.upload(transcribe_audio, audio_input, transcribed)

    clear_btn = gr.Button("Clear All")
    clear_btn.click(lambda: ([], "", None), None, [chatbot, text_input, audio_output])
    #gr.Markdown("---")
    gr.Markdown(load_readme())

demo.queue().launch()