Update app.py

app.py

@@ -3,28 +3,36 @@
 # ===================================================================================
 from __future__ import annotations
 import os
+import sys
 import base64
 import struct
 import textwrap
 import requests
 import atexit
-import tempfile
 from typing import List, Dict, Tuple, Generator
 
 # --- 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")
+os.environ.setdefault("GRADIO_ANALYTICS_ENABLED", "false")  # truly disable analytics
+os.environ.setdefault("TORCHAUDIO_USE_FFMPEG", "0")  # avoid torchaudio ffmpeg path entirely
 
-# --- Load .env early (HF_TOKEN / SECRET_TOKEN) ---
+# --- .env early (HF_TOKEN / SECRET_TOKEN) ---
 from dotenv import load_dotenv
 load_dotenv()
 
-# --- Hugging Face Spaces & ZeroGPU ---
+# --- NumPy sanity (Torch 2.2.2 expects NumPy 1.x in your stack) ---
+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) for this Space."
+    )
+
+# --- Hugging Face Spaces & ZeroGPU (import BEFORE CUDA libs) ---
 try:
-    import spaces
-except ImportError:
+    import spaces  # Required for ZeroGPU on HF
+except Exception:
     class _SpacesShim:
         def GPU(self, *args, **kwargs):
             def _wrap(fn):
@@ -34,19 +42,20 @@ except ImportError:
 
 import gradio as gr
 
-# --- Core ML & Data Libraries ---
+# --- 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
-import torchaudio
-import soundfile as sf
+
+# --- Audio decoding (we'll use ffmpeg-python to avoid torchaudio/torio) ---
+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
-from TTS.utils.generic_utils import get_user_data_dir
+import TTS.tts.models.xtts as xtts_module  # for monkey-patching load_audio
 
 # --- Text & Audio Processing ---
 import nltk
@@ -58,12 +67,15 @@ import noisereduce as nr
 # 2) GLOBALS & HELPERS
 # ===================================================================================
 
+# NLTK data
 nltk.download("punkt", quiet=True)
 
+# Cached models & latents
 tts_model: Xtts | None = None
 llm_model: Llama | None = None
-voice_latents: Dict[str, Tuple[torch.Tensor, torch.Tensor]] = {}
+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"
@@ -71,6 +83,7 @@ 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')."
@@ -83,6 +96,7 @@ ROLE_PROMPTS["Pirate"] = (
     "Keep answers short, as if in a real conversation. Only provide the words AI Beard would speak."
 )
 
+# ---------- small utils ----------
 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
@@ -115,11 +129,44 @@ def format_prompt_zephyr(message: str, history: List[Tuple[str, str | None]], sy
     prompt += f"<|user|>\n{message}</s><|assistant|>"
     return prompt
 
+# ---------- robust audio decode (24k mono via ffmpeg) ----------
+def _decode_audio_ffmpeg_to_24k_mono(path: str) -> Tuple[np.ndarray, int]:
+    """Return float32 waveform in [-1,1], 24 kHz mono."""
+    try:
+        out, _ = (
+            ffmpeg
+            .input(path)
+            .output("pipe:", format="s16le", acodec="pcm_s16le", ac=1, ar=24000)
+            .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).copy()
+        return wav, 24000
+    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):
+    wav, sr = _decode_audio_ffmpeg_to_24k_mono(audiopath)
+    # XTTS expects (audio, sr) and will handle truncation/conditioning windows.
+    return wav, sr
+
+xtts_module.load_audio = _patched_load_audio  # <- critical fix
+
+# ---------- where Coqui caches models (avoid get_user_data_dir import) ----------
+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 (RUN BEFORE FIRST INFERENCE)
 # ===================================================================================
 
 def precache_assets() -> None:
+    """Download voice WAVs, XTTS weights, and Zephyr GGUF to local cache before any inference."""
+    # Voices
     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/"
@@ -135,35 +182,43 @@ def precache_assets() -> None:
             except Exception as e:
                 print(f"Failed to download {name}: {e}")
 
+    # XTTS model files
     print("Pre-caching XTTS v2 model files...")
     ModelManager().download_model("tts_models/multilingual/multi-dataset/xtts_v2")
 
+    # LLM GGUF
     print("Pre-caching Zephyr GGUF...")
     try:
         hf_hub_download(
             repo_id="TheBloke/zephyr-7B-beta-GGUF",
             filename="zephyr-7b-beta.Q5_K_M.gguf",
-            local_dir_use_symlinks=False,
+            force_download=False
         )
     except Exception as e:
         print(f"Warning: GGUF pre-cache error: {e}")
 
 def _load_xtts(device: str) -> Xtts:
+    """Load XTTS from the local cache. Use checkpoint_dir to avoid None path bugs."""
     print("Loading Coqui XTTS V2 model (CPU first)...")
     model_name = "tts_models/multilingual/multi-dataset/xtts_v2"
-    model_dir = os.path.join(get_user_data_dir("tts"), model_name.replace("/", "--"))
-    if not os.path.exists(model_dir):
-        ModelManager().download_model(model_name)
+    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.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) on CPU so it's ready immediately."""
     print("Loading LLM (Zephyr GGUF) on CPU...")
     zephyr_model_path = hf_hub_download(
         repo_id="TheBloke/zephyr-7B-beta-GGUF",
@@ -171,126 +226,143 @@ def _load_llama() -> Llama:
     )
     llm = Llama(
         model_path=zephyr_model_path,
-        n_gpu_layers=0, n_ctx=4096, n_batch=512, verbose=False
+        n_gpu_layers=0,   # CPU by default
+        n_ctx=4096,
+        n_batch=512,
+        verbose=False
     )
     print("LLM loaded (CPU).")
     return llm
 
-def load_and_resample_audio(path: str, target_sr: int = 24000) -> torch.Tensor:
-    """Loads audio, converts to a Torch tensor, and resamples if needed."""
-    try:
-        audio_np, original_sr = sf.read(path, dtype='float32')
-        if audio_np.ndim > 1:
-            audio_np = np.mean(audio_np, axis=1)
-        waveform = torch.from_numpy(audio_np).float()
-        
-        if original_sr != target_sr:
-            print(f"Resampling audio from {original_sr}Hz to {target_sr}Hz.")
-            resampler = torchaudio.transforms.Resample(orig_freq=original_sr, new_freq=target_sr)
-            waveform = resampler(waveform)
-            
-        return waveform.unsqueeze(0)
-    except Exception as e:
-        print(f"Error loading audio file {path}: {e}")
-        raise
-
 def init_models_and_latents() -> None:
-    """Preload models and compute voice latents, using temporary files for compatibility."""
+    """Preload TTS and LLM on CPU and compute voice latents once (using patched audio loader)."""
     global tts_model, llm_model, voice_latents
 
     if tts_model is None:
-        tts_model = _load_xtts(device="cpu")
+        tts_model = _load_xtts(device="cpu")  # keep on CPU at startup
 
     if llm_model is None:
         llm_model = _load_llama()
 
+    # Pre-compute latents once (CPU OK); uses patched loader (ffmpeg) under the hood
     if not voice_latents:
         print("Computing voice conditioning latents...")
-        with tempfile.TemporaryDirectory() as temp_dir:
-            voice_files = {
-                "Cloée": "cloee-1.wav", "Julian": "julian-bedtime-style-1.wav",
-                "Pirate": "pirate_by_coqui.wav", "Thera": "thera-1.wav",
-            }
-            for role, filename in voice_files.items():
-                original_path = os.path.join("voices", filename)
-                resampled_waveform = load_and_resample_audio(original_path)
-                
-                temp_path = os.path.join(temp_dir, f"resampled_{filename}")
-                
-                # --- FIX: Replace torchaudio.save with the more stable soundfile.write ---
-                numpy_waveform = resampled_waveform.squeeze(0).cpu().numpy()
-                sf.write(temp_path, numpy_waveform, 24000)
-                
-                voice_latents[role] = tts_model.get_conditioning_latents(
-                    audio_path=temp_path,
-                    gpt_cond_len=30,
-                    max_ref_length=60
-                )
+        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)
+            voice_latents[role] = tts_model.get_conditioning_latents(
+                audio_path=path, gpt_cond_len=30, max_ref_length=60
+            )
         print("Voice latents ready.")
 
+# Ensure we close Llama cleanly to avoid __del__ issues at interpreter shutdown
 def _close_llm():
     global llm_model
-    if llm_model is not None:
-        del 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, sys_prompt: str) -> Generator[str, None, None]:
-    formatted_prompt = format_prompt_zephyr(prompt, history, sys_prompt)
+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
+        formatted_prompt,
+        temperature=0.7,
+        max_tokens=512,
+        top_p=0.95,
+        stop=LLM_STOP_WORDS,
+        stream=True
     )
     for response in stream:
-        yield response["choices"][0]["text"]
-
-def generate_audio_stream(tts_instance: Xtts, text: str, lang: str, latents: Tuple) -> Generator[bytes, None, None]:
+        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 generate_audio_stream(tts_instance: Xtts, text: str, language: str,
+                          latents: Tuple[np.ndarray, np.ndarray]) -> Generator[bytes, None, None]:
     gpt_cond_latent, speaker_embedding = latents
-    for chunk in tts_instance.inference_stream(
-        text, lang, gpt_cond_latent, speaker_embedding, temperature=0.85,
-    ):
-        if chunk is not None:
-            yield chunk.detach().cpu().numpy().squeeze().tobytes()
+    try:
+        for chunk in tts_instance.inference_stream(
+            text=text,
+            language=language,
+            gpt_cond_latent=gpt_cond_latent,
+            speaker_embedding=speaker_embedding,
+            temperature=0.85,
+        ):
+            if chunk is not None:
+                yield chunk.detach().cpu().numpy().squeeze().tobytes()
+    except RuntimeError as e:
+        print(f"Error during TTS inference: {e}")
+        if "device-side assert" in str(e) and api:
+            try:
+                gr.Warning("Critical GPU error. Attempting to restart the Space...")
+                api.restart_space(repo_id=repo_id)
+            except Exception:
+                pass
 
 # ===================================================================================
 # 5) ZERO-GPU ENTRYPOINT
 # ===================================================================================
 
-@spaces.GPU(duration=120)
+@spaces.GPU(duration=120)  # Request GPU for up to 120s; adjust as needed
 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 []
 
-    if tts_model is None or llm_model is None:
-        raise gr.Error("Models not initialized. Please restart the Space.")
+    # Ensure models/latents exist
+    if tts_model is None or llm_model is None or not voice_latents:
+        init_models_and_latents()
 
+    # Move XTTS to CUDA for this call if GPU is 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 on CPU)
+    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 []
 
-        history: List[Tuple[str, str | None]] = [(input_text, None)]
-        full_story_text = "".join(
-            generate_text_stream(llm_model, history[-1][0], history[:-1], 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"
 
-        sentences = split_sentences(full_story_text, SENTENCE_SPLIT_LENGTH)
-        lang = langid.classify(sentences[0])[0] if sentences else "en"
-        results: List[Dict[str, str]] = []
+    results: List[Dict[str, str]] = []
+    for sentence in sentences:
+        if not any(c.isalnum() for c in sentence):
+            continue
 
-        for sentence in sentences:
-            if not any(c.isalnum() for c in sentence):
-                continue
-            
-            audio_chunks = generate_audio_stream(tts_model, sentence, lang, voice_latents[chatbot_role])
-            pcm_data = b"".join(chunk for chunk in audio_chunks if chunk)
+        audio_chunks = generate_audio_stream(tts_model, sentence, lang, voice_latents[chatbot_role])
+        pcm_data = b"".join(chunk for chunk in audio_chunks if chunk)
 
+        # 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
@@ -298,45 +370,43 @@ def generate_story_and_speech(secret_token_input: str, input_text: str, chatbot_
                 final_pcm = (reduced * 32767).astype(np.int16).tobytes()
             else:
                 final_pcm = pcm_data
-            
-            b64_wav = base64.b64encode(pcm_to_wav(final_pcm)).decode("utf-8")
-            results.append({"text": sentence, "audio": b64_wav})
-
-        return results
-    
-    finally:
-        if tts_model is not None:
-            tts_model.to("cpu")
+        except Exception:
+            final_pcm = pcm_data
+
+        b64_wav = base64.b64encode(pcm_to_wav(final_pcm)).decode("utf-8")
+        results.append({"text": sentence, "audio": b64_wav})
+
+    # Release GPU immediately
+    try:
+        tts_model.to("cpu")
+    except Exception:
+        pass
+
+    return results
 
 # ===================================================================================
 # 6) STARTUP: PRECACHE & UI
 # ===================================================================================
 
-def build_ui() -> gr.Blocks:
-    with gr.Blocks() as demo:
-        gr.Markdown("# AI Storyteller with ZeroGPU")
-        gr.Markdown("Enter a prompt to generate a short story with voice narration using on-demand GPU.")
-        
-        with gr.Row():
-            secret_token = gr.Textbox(label="Secret Token", type="password", value=SECRET_TOKEN)
-            storyteller = gr.Dropdown(choices=ROLES, label="Select a Storyteller", value="Cloée")
-        
-        prompt = gr.Textbox(placeholder="What should the story be about?", label="Story Prompt")
-        output = gr.JSON(label="Story and Audio Output")
-        
-        prompt.submit(
-            fn=generate_story_and_speech,
-            inputs=[secret_token, prompt, storyteller],
-            outputs=output,
-        )
-
-    return demo
+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 using on-demand GPU.",
+        flagging_mode="never",
+    )
 
 if __name__ == "__main__":
     print("===== Startup: pre-cache assets and preload models =====")
-    precache_assets()
-    init_models_and_latents()
+    precache_assets()              # 1) download everything to disk
+    init_models_and_latents()      # 2) load models on CPU + compute voice latents (via patched loader)
     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")))
+    demo.queue().launch(server_name="0.0.0.0", server_port=int(os.getenv("PORT", "7860")))