app.py

# ===================================================================================
# 1) SETUP & IMPORTS
# ===================================================================================
from __future__ import annotations

import os
import sys
import re
import base64
import struct
import textwrap
import requests
import atexit
import inspect
from typing import List, Dict, Tuple, Generator, Any

# --- Fast, safe defaults ---
os.environ.setdefault("HF_HUB_ENABLE_HF_TRANSFER", "1")
os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
os.environ.setdefault("COQUI_TOS_AGREED", "1")
os.environ.setdefault("GRADIO_ANALYTICS_ENABLED", "false")   # truly disable analytics
os.environ.setdefault("TORCHAUDIO_USE_FFMPEG", "0")          # avoid torchaudio/ffmpeg linkage quirks

# --- .env early (HF_TOKEN / SECRET_TOKEN) ---
from dotenv import load_dotenv
load_dotenv()

# --- NumPy sanity (Torch 2.2.x prefers NumPy 1.x) ---
import numpy as _np
if int(_np.__version__.split(".", 1)[0]) >= 2:
    raise RuntimeError(
        f"Detected numpy=={_np.__version__}. Please ensure numpy<2 (e.g., 1.26.4)."
    )

# --- Pandas compat shims (Gradio & mixed versions) ---
try:
    import pandas as pd
    from pandas._config.config import register_option
    # Option shim
    try:
        pd.get_option("future.no_silent_downcasting")
    except Exception:
        register_option("future.no_silent_downcasting", False, validator=None, doc="compat shim for Gradio")
    # infer_objects(copy=...) shim (older pandas lacks 'copy' kwarg)
    if hasattr(pd, "DataFrame"):
        try:
            sig = inspect.signature(pd.DataFrame.infer_objects)
            if "copy" not in sig.parameters:
                _orig_infer_objects = pd.DataFrame.infer_objects
                def _infer_objects_compat(self, *args, **kwargs):
                    kwargs.pop("copy", None)
                    return _orig_infer_objects(self, *args, **kwargs)
                pd.DataFrame.infer_objects = _infer_objects_compat
        except Exception:
            pass
except Exception:
    pd = None  # ok if pandas is unavailable

# --- Hugging Face Spaces & ZeroGPU (import BEFORE CUDA libs) ---
try:
    import spaces  # Required for ZeroGPU on HF
except Exception:
    class _SpacesShim:
        def GPU(self, *args, **kwargs):
            def _wrap(fn):
                return fn
            return _wrap
    spaces = _SpacesShim()

import gradio as gr

# --- Core ML & Data Libraries (after spaces import) ---
import torch
import numpy as np
from huggingface_hub import HfApi, hf_hub_download
from llama_cpp import Llama

# --- Audio decoding (pure ffmpeg-python; no torchaudio) ---
import ffmpeg

# --- TTS Libraries ---
from TTS.tts.configs.xtts_config import XttsConfig
from TTS.tts.models.xtts import Xtts
from TTS.utils.manage import ModelManager
import TTS.tts.models.xtts as xtts_module  # for monkey-patching load_audio

# --- Text & Audio Processing ---
import nltk
import langid
import emoji
import noisereduce as nr


# ===================================================================================
# 2) GLOBALS & HELPERS
# ===================================================================================

def _ensure_nltk() -> None:
    # Newer NLTK splits data into 'punkt' and 'punkt_tab'
    for pkg in ("punkt", "punkt_tab"):
        try:
            if pkg == "punkt":
                nltk.data.find("tokenizers/punkt")
            else:
                nltk.data.find("tokenizers/punkt_tab")
        except LookupError:
            nltk.download(pkg, quiet=True)

_ensure_nltk()

# Models & caches
tts_model: Xtts | None = None
llm_model: Llama | None = None

# Store latents as NumPy on CPU for portability; convert to device at inference time
voice_latents: Dict[str, Tuple[np.ndarray, np.ndarray]] = {}

# Config
HF_TOKEN = os.environ.get("HF_TOKEN")
api = HfApi(token=HF_TOKEN) if HF_TOKEN else None
repo_id = "ruslanmv/ai-story-server"
SECRET_TOKEN = os.getenv("SECRET_TOKEN", "secret")
SENTENCE_SPLIT_LENGTH = 250
LLM_STOP_WORDS = ["</s>", "<|user|>", "/s>"]

# System prompts and roles
default_system_message = (
    "You're a storyteller crafting a short tale for young listeners. Keep sentences short and simple. "
    "Use narrative style only, without lists or complex words. Type numbers as words (e.g., 'ten')."
)
system_message = os.environ.get("SYSTEM_MESSAGE", default_system_message)
ROLES = ["Cloée", "Julian", "Pirate", "Thera"]
ROLE_PROMPTS = {role: system_message for role in ROLES}
ROLE_PROMPTS["Pirate"] = (
    "You are AI Beard, a pirate. Craft your response from his first-person perspective. "
    "Keep answers short, as if in a real conversation. Only provide the words AI Beard would speak."
)


# ---------- tiny utilities ----------
def _model_device(m: torch.nn.Module) -> torch.device:
    try:
        return next(m.parameters()).device
    except StopIteration:
        return torch.device("cpu")

def _to_device_float_tensor(x: Any, device: torch.device) -> torch.Tensor:
    if isinstance(x, np.ndarray):
        return torch.from_numpy(x).float().to(device)
    if torch.is_tensor(x):
        return x.to(device, dtype=torch.float32)
    return torch.as_tensor(x, dtype=torch.float32, device=device)

def _latents_for_device(latents: Tuple[Any, Any], device: torch.device) -> Tuple[torch.Tensor, torch.Tensor]:
    gpt_cond, spk = latents
    return _to_device_float_tensor(gpt_cond, device), _to_device_float_tensor(spk, device)

def pcm_to_wav(pcm_data: bytes, sample_rate: int = 24000, channels: int = 1, bit_depth: int = 16) -> bytes:
    if pcm_data.startswith(b"RIFF"):
        return pcm_data
    byte_rate = sample_rate * channels * bit_depth // 8
    block_align = channels * bit_depth // 8
    chunk_size = 36 + len(pcm_data)
    header = struct.pack(
        "<4sI4s4sIHHIIHH4sI",
        b"RIFF", chunk_size, b"WAVE", b"fmt ",
        16, 1, channels, sample_rate, byte_rate, block_align, bit_depth,
        b"data", len(pcm_data)
    )
    return header + pcm_data

def split_sentences(text: str, max_len: int) -> List[str]:
    # Try NLTK; if it fails for any reason, fallback to a simple regex splitter.
    try:
        sentences = nltk.sent_tokenize(text)
    except Exception:
        sentences = re.split(r"(?<=[\.\!\?])\s+", text)
    chunks: List[str] = []
    for sent in sentences:
        if len(sent) > max_len:
            chunks.extend(textwrap.wrap(sent, max_len, break_long_words=True))
        else:
            if sent:
                chunks.append(sent)
    return chunks

def format_prompt_zephyr(message: str, history: List[Tuple[str, str | None]], system_message: str) -> str:
    prompt = f"<|system|>\n{system_message}</s>"
    for user_prompt, bot_response in history:
        if bot_response:
            prompt += f"<|user|>\n{user_prompt}</s><|assistant|>\n{bot_response}</s>"
    prompt += f"<|user|>\n{message}</s><|assistant|>"
    return prompt


# ---------- robust audio decode (mono via ffmpeg) ----------
def _decode_audio_ffmpeg_to_mono(path: str, target_sr: int) -> np.ndarray:
    """
    Return float32 waveform in [-1, 1], mono, resampled to target_sr.
    Shape: (samples,)
    """
    try:
        out, _ = (
            ffmpeg
            .input(path)
            .output("pipe:", format="s16le", acodec="pcm_s16le", ac=1, ar=target_sr)
            .run(capture_stdout=True, capture_stderr=True, cmd="ffmpeg")
        )
        pcm = np.frombuffer(out, dtype=np.int16)
        if pcm.size == 0:
            raise RuntimeError("ffmpeg produced empty audio.")
        wav = (pcm.astype(np.float32) / 32767.0)
        return wav
    except ffmpeg.Error as e:
        raise RuntimeError(f"ffmpeg decode failed: {e.stderr.decode(errors='ignore') if e.stderr else e}") from e


# ---------- monkey-patch XTTS internal loader to avoid torchaudio/torio ----------
def _patched_load_audio(audiopath: str, load_sr: int):
    """
    Match XTTS' expected return type:
    - returns a torch.FloatTensor shaped [1, samples], normalized to [-1, 1],
      already resampled to `load_sr`.
    - DO NOT return (audio, sr) tuple.
    """
    wav = _decode_audio_ffmpeg_to_mono(audiopath, target_sr=load_sr)
    import torch as _torch  # local import to avoid circularities
    audio = _torch.from_numpy(wav).float().unsqueeze(0)  # [1, N] on CPU
    return audio

xtts_module.load_audio = _patched_load_audio
try:
    import TTS.utils.audio as _tts_audio_mod
    _tts_audio_mod.load_audio = _patched_load_audio
except Exception:
    pass


def _coqui_cache_dir() -> str:
    # Matches what TTS uses on Linux: ~/.local/share/tts
    return os.path.join(os.path.expanduser("~"), ".local", "share", "tts")


# ===================================================================================
# 3) PRECACHE & MODEL LOADERS (CPU at startup to avoid ZeroGPU issues)
# ===================================================================================

def precache_assets() -> None:
    """Download voice WAVs, XTTS weights, and Zephyr GGUF to local cache before any inference."""
    print("Pre-caching voice files...")
    file_names = ["cloee-1.wav", "julian-bedtime-style-1.wav", "pirate_by_coqui.wav", "thera-1.wav"]
    base_url = "https://raw.githubusercontent.com/ruslanmv/ai-story-server/main/voices/"
    os.makedirs("voices", exist_ok=True)
    for name in file_names:
        dst = os.path.join("voices", name)
        if not os.path.exists(dst):
            try:
                resp = requests.get(base_url + name, timeout=30)
                resp.raise_for_status()
                with open(dst, "wb") as f:
                    f.write(resp.content)
            except Exception as e:
                print(f"Failed to download {name}: {e}")

    print("Pre-caching XTTS v2 model files...")
    ModelManager().download_model("tts_models/multilingual/multi-dataset/xtts_v2")

    print("Pre-caching Zephyr GGUF...")
    try:
        hf_hub_download(
            repo_id="TheBloke/zephyr-7B-beta-GGUF",
            filename="zephyr-7b-beta.Q5_K_M.gguf",
            force_download=False
        )
    except Exception as e:
        print(f"Warning: GGUF pre-cache error: {e}")


def _load_xtts(device: str = "cpu") -> Xtts:
    """Load XTTS from the local cache. Keep CPU at startup to avoid ZeroGPU device mixups."""
    print(f"Loading Coqui XTTS V2 model on {device.upper()}...")
    model_name = "tts_models/multilingual/multi-dataset/xtts_v2"
    ModelManager().download_model(model_name)  # idempotent
    model_dir = os.path.join(_coqui_cache_dir(), model_name.replace("/", "--"))

    cfg = XttsConfig()
    cfg.load_json(os.path.join(model_dir, "config.json"))
    model = Xtts.init_from_config(cfg)
    model.load_checkpoint(
        cfg,
        checkpoint_dir=model_dir,
        eval=True,
        use_deepspeed=False,
    )
    model.to(device)
    print("XTTS model loaded.")
    return model


def _load_llama() -> Llama:
    """
    Load Llama (Zephyr GGUF).
    Keep simple & robust: default to CPU (works everywhere).
    """
    print("Loading LLM (Zephyr GGUF)...")
    zephyr_model_path = hf_hub_download(
        repo_id="TheBloke/zephyr-7B-beta-GGUF",
        filename="zephyr-7b-beta.Q5_K_M.gguf"
    )
    llm = Llama(
        model_path=zephyr_model_path,
        n_gpu_layers=0,     # CPU-only for reliability across Spaces/ZeroGPU
        n_ctx=4096,
        n_batch=512,
        verbose=False
    )
    print("LLM loaded (CPU).")
    return llm


def init_models_and_latents() -> None:
    """
    Preload models on CPU and compute voice latents on CPU.
    This avoids ZeroGPU's "mixed device" errors from torchaudio-based resampling.
    """
    global tts_model, llm_model, voice_latents

    if tts_model is None:
        tts_model = _load_xtts(device="cpu")  # always CPU at startup

    if llm_model is None:
        llm_model = _load_llama()

    if not voice_latents:
        print("Computing voice conditioning latents (CPU)...")
        # Ensure the TTS model is on CPU while computing latents
        orig_dev = _model_device(tts_model)
        if orig_dev.type != "cpu":
            tts_model.to("cpu")

        with torch.no_grad():
            for role, filename in [
                ("Cloée", "cloee-1.wav"),
                ("Julian", "julian-bedtime-style-1.wav"),
                ("Pirate", "pirate_by_coqui.wav"),
                ("Thera", "thera-1.wav"),
            ]:
                path = os.path.join("voices", filename)
                gpt_lat, spk_emb = tts_model.get_conditioning_latents(
                    audio_path=path, gpt_cond_len=30, max_ref_length=60
                )
                # Store as NumPy on CPU; convert to device on demand later
                voice_latents[role] = (
                    gpt_lat.detach().cpu().numpy(),
                    spk_emb.detach().cpu().numpy(),
                )

        # Return model to original device (keep CPU at startup for safety)
        if orig_dev.type != "cpu":
            tts_model.to(orig_dev)

        print("Voice latents ready.")


# Ensure we close Llama cleanly to avoid __del__ issues at interpreter shutdown
def _close_llm():
    global llm_model
    try:
        if llm_model is not None:
            llm_model.close()
    except Exception:
        pass
atexit.register(_close_llm)


# ===================================================================================
# 4) INFERENCE HELPERS
# ===================================================================================

def generate_text_stream(llm_instance: Llama, prompt: str,
                         history: List[Tuple[str, str | None]],
                         system_message_text: str) -> Generator[str, None, None]:
    formatted_prompt = format_prompt_zephyr(prompt, history, system_message_text)
    stream = llm_instance(
        formatted_prompt,
        temperature=0.7,
        max_tokens=512,
        top_p=0.95,
        stop=LLM_STOP_WORDS,
        stream=True
    )
    for response in stream:
        ch = response["choices"][0]["text"]
        try:
            is_single_emoji = (len(ch) == 1 and emoji.is_emoji(ch))
        except Exception:
            is_single_emoji = False
        if "<|user|>" in ch or is_single_emoji:
            continue
        yield ch


def _extract_waveform_to_numpy(wav_any: Any) -> np.ndarray:
    """
    Normalize various XTTS inference() return shapes/types to 1-D float32 numpy in [-1, 1].
    """
    if isinstance(wav_any, dict):
        for k in ("wav", "audio", "samples", "waveform"):
            if k in wav_any:
                wav_any = wav_any[k]
                break
    if torch.is_tensor(wav_any):
        arr = wav_any.detach().cpu().numpy()
    else:
        arr = np.asarray(wav_any)
    arr = np.squeeze(arr).astype(np.float32)
    # If not already normalized, attempt to scale if max > 1 (heuristic)
    maxabs = np.max(np.abs(arr)) if arr.size else 1.0
    if maxabs > 1.5:  # likely int16 or higher-amplitude float
        arr = arr / 32767.0
    arr = np.clip(arr, -1.0, 1.0)
    return arr


def synthesize_sentence_pcm16(tts_instance: Xtts, text: str, language: str,
                              latents: Tuple[np.ndarray, np.ndarray]) -> bytes:
    """
    Use non-streaming XTTS inference() to avoid GPT2InferenceModel streaming bug.
    Returns PCM16 bytes at 24 kHz mono.
    """
    device = _model_device(tts_instance)
    gpt_cond_latent_t, speaker_embedding_t = _latents_for_device(latents, device)

    with torch.no_grad():
        out = tts_instance.inference(
            text=text,
            language=language,
            gpt_cond_latent=gpt_cond_latent_t,
            speaker_embedding=speaker_embedding_t,
            temperature=0.85,
        )

    f32 = _extract_waveform_to_numpy(out)
    s16 = (np.clip(f32, -1.0, 1.0) * 32767.0).astype(np.int16)
    return s16.tobytes()


# ===================================================================================
# 5) ZERO-GPU ENTRYPOINT (safe on native GPU as well)
# ===================================================================================

@spaces.GPU(duration=120)  # GPU ops happen inside when on ZeroGPU
def generate_story_and_speech(secret_token_input: str, input_text: str, chatbot_role: str) -> List[Dict[str, str]]:
    if secret_token_input != SECRET_TOKEN:
        raise gr.Error("Invalid secret token provided.")
    if not input_text:
        return []

    # Ensure models/latents exist (loaded on CPU)
    if tts_model is None or llm_model is None or not voice_latents:
        init_models_and_latents()

    # During the GPU window, move XTTS to CUDA if available; otherwise stay on CPU
    try:
        if torch.cuda.is_available():
            tts_model.to("cuda")
        else:
            tts_model.to("cpu")
    except Exception:
        tts_model.to("cpu")

    # Generate story text (LLM kept CPU for simplicity & reliability)
    history: List[Tuple[str, str | None]] = [(input_text, None)]
    full_story_text = "".join(
        generate_text_stream(llm_model, history[-1][0], history[:-1], system_message_text=ROLE_PROMPTS[chatbot_role])
    ).strip()
    if not full_story_text:
        return []

    # Split into TTS-friendly sentences
    sentences = split_sentences(full_story_text, SENTENCE_SPLIT_LENGTH)
    lang = langid.classify(sentences[0])[0] if sentences else "en"

    results: List[Dict[str, str]] = []
    for sentence in sentences:
        if not any(c.isalnum() for c in sentence):
            continue

        # Synthesize whole sentence (non-stream) to avoid streaming bug
        pcm_data = synthesize_sentence_pcm16(tts_model, sentence, lang, voice_latents[chatbot_role])

        # Optional noise reduction (best-effort)
        try:
            data_s16 = np.frombuffer(pcm_data, dtype=np.int16)
            if data_s16.size > 0:
                float_data = (data_s16.astype(np.float32) / 32767.0)
                reduced = nr.reduce_noise(y=float_data, sr=24000)
                final_pcm = np.clip(reduced * 32767.0, -32768, 32767).astype(np.int16).tobytes()
            else:
                final_pcm = pcm_data
        except Exception:
            final_pcm = pcm_data

        b64_wav = base64.b64encode(pcm_to_wav(final_pcm, sample_rate=24000, channels=1, bit_depth=16)).decode("utf-8")
        results.append({"text": sentence, "audio": b64_wav})

    # Leave model on CPU after the ZeroGPU window
    try:
        tts_model.to("cpu")
    except Exception:
        pass

    return results


# ===================================================================================
# 6) STARTUP: PRECACHE & UI
# ===================================================================================

def build_ui() -> gr.Interface:
    return gr.Interface(
        fn=generate_story_and_speech,
        inputs=[
            gr.Textbox(label="Secret Token", type="password", value=SECRET_TOKEN),
            gr.Textbox(placeholder="What should the story be about?", label="Story Prompt"),
            gr.Dropdown(choices=ROLES, label="Select a Storyteller", value="Cloée"),
        ],
        outputs=gr.JSON(label="Story and Audio Output"),
        title="AI Storyteller with ZeroGPU",
        description="Enter a prompt to generate a short story with voice narration. Uses GPU only within the generation call when available.",
        flagging_mode="never",
    )

if __name__ == "__main__":
    print("===== Startup: pre-cache assets and preload models (CPU) =====")
    print(f"Python: {sys.version.split()[0]} | Torch CUDA available: {torch.cuda.is_available()}")
    precache_assets()              # 1) download everything to disk
    init_models_and_latents()      # 2) load models on CPU + compute voice latents on CPU
    print("Models and assets ready. Launching UI...")

    demo = build_ui()
    demo.queue().launch(
        server_name="0.0.0.0",
        server_port=int(os.getenv("PORT", "7860")),
        ssr_mode=False,  # disable experimental SSR noise
    )