Update app.py

app.py

@@ -1,7 +1,9 @@
 import os
-os.environ["CUDA_VISIBLE_DEVICES"] = ""  # force CPU-only
+# Paksa PyTorch agar hanya melihat CPU dan matikan CUDA murni sebelum library lain di-import
+os.environ["CUDA_VISIBLE_DEVICES"] = ""
 
 import re
+import gc
 import inspect
 import tempfile
 import traceback
@@ -12,23 +14,17 @@ import torch
 import torchaudio as ta
 import gradio as gr
 
-# =========================
-# CONFIG (ANTI NGARET)
-# =========================
-MODEL_REPO = "grandhigh/Chatterbox-TTS-Indonesian"
-CHECKPOINT_FILENAME = "t3_cfg.safetensors"
-DEVICE = "cpu"
+# =====================================================================
+# OPTIMASI EKSTREM LEVEL CPU & PYTORCH (ANTI-NGARET & DEPREKASI)
+# =====================================================================
+# Batasi thread PyTorch secara agresif agar tidak rebutan core di shared CPU
+torch.set_num_threads(2)
+torch.set_num_interop_threads(2)
 
-# Batasi beban CPU
-MAX_TOTAL_CHARS = int(os.getenv("MAX_TOTAL_CHARS", "2400"))       # total karakter per request
-MAX_CHARS_PER_CHUNK = int(os.getenv("MAX_CHARS_PER_CHUNK", "220"))# karakter per chunk
-MAX_CHUNKS = int(os.getenv("MAX_CHUNKS", "12"))                   # maksimal jumlah chunk
-PAUSE_SECONDS = float(os.getenv("PAUSE_SECONDS", "0.15"))         # jeda antar chunk
-DOWNLOAD_TIMEOUT = int(os.getenv("DOWNLOAD_TIMEOUT", "90"))
+# Matikan kalkulasi gradient secara global karena ini murni inference/synthesis
+torch.set_grad_enabled(False)
 
-# =========================
-# HARD PATCH CPU DESERIALIZE
-# =========================
+# Hard patch untuk mendepresiasi CUDA di library pihak ketiga
 torch.cuda.is_available = lambda: False  # noqa: E731
 
 _original_torch_load = torch.load
@@ -44,9 +40,22 @@ if hasattr(torch.jit, "load"):
         return _original_jit_load(*args, **kwargs)
     torch.jit.load = _jit_load_cpu
 
-# =========================
-# MODEL IMPORT
-# =========================
+# =====================================================================
+# CONFIG (BATASAN REASONS & KONTROL MEMORI)
+# =====================================================================
+MODEL_REPO = "grandhigh/Chatterbox-TTS-Indonesian"
+CHECKPOINT_FILENAME = "t3_cfg.safetensors"
+DEVICE = "cpu"
+
+MAX_TOTAL_CHARS = int(os.getenv("MAX_TOTAL_CHARS", "2400"))       # total karakter per request
+MAX_CHARS_PER_CHUNK = int(os.getenv("MAX_CHARS_PER_CHUNK", "220")) # karakter per chunk
+MAX_CHUNKS = int(os.getenv("MAX_CHUNKS", "12"))                     # maksimal jumlah chunk
+PAUSE_SECONDS = float(os.getenv("PAUSE_SECONDS", "0.15"))          # jeda antar chunk
+DOWNLOAD_TIMEOUT = int(os.getenv("DOWNLOAD_TIMEOUT", "90"))
+
+# =====================================================================
+# MODEL IMPORT & SINGLETON LAZY LOADING
+# =====================================================================
 from chatterbox.tts import ChatterboxTTS
 from huggingface_hub import hf_hub_download
 from safetensors.torch import load_file
@@ -89,17 +98,14 @@ def _download_wav(url: str) -> str:
 
 
 def _resolve_audio_input(audio_file, audio_url: str):
-    # gr.Audio(type="filepath") -> string path
     if isinstance(audio_file, str) and audio_file.strip():
         return audio_file
 
-    # fallback dict
     if isinstance(audio_file, dict):
         p = audio_file.get("path")
         if p:
             return p
 
-    # URL fallback
     if audio_url and audio_url.strip():
         return _download_wav(audio_url.strip())
 
@@ -120,9 +126,7 @@ def _split_text_safely(text: str, max_chars: int = MAX_CHARS_PER_CHUNK):
     if not text:
         return []
 
-    # Split kalimat
     sentences = re.split(r"(?<=[.!?])\s+", text)
-
     chunks = []
     current = ""
 
@@ -131,7 +135,6 @@ def _split_text_safely(text: str, max_chars: int = MAX_CHARS_PER_CHUNK):
         if not s:
             continue
 
-        # Jika kalimat panjang, pecah pakai koma/titik koma/titik dua
         parts = [s] if len(s) <= max_chars else re.split(r"(?<=[,;:])\s+", s)
 
         for p in parts:
@@ -139,7 +142,6 @@ def _split_text_safely(text: str, max_chars: int = MAX_CHARS_PER_CHUNK):
             if not p:
                 continue
 
-            # kalau masih kepanjangan, hard-cut berbasis kata
             if len(p) > max_chars:
                 words = p.split()
                 tmp = ""
@@ -174,11 +176,9 @@ def _generate_with_safe_kwargs(model, text: str, prompt_path: str):
     params = sig.parameters
     kwargs = {}
 
-    # prompt audio
     if "audio_prompt_path" in params:
         kwargs["audio_prompt_path"] = prompt_path
 
-    # Stabilitas & kecepatan (kalau param tersedia)
     if "temperature" in params:
         kwargs["temperature"] = 0.05
     if "top_p" in params:
@@ -188,9 +188,8 @@ def _generate_with_safe_kwargs(model, text: str, prompt_path: str):
     if "cfg_weight" in params:
         kwargs["cfg_weight"] = 0.3
     if "max_new_tokens" in params:
-        kwargs["max_new_tokens"] = 260  # cegah runaway generation
+        kwargs["max_new_tokens"] = 260  # Kunci agar model tidak runaway loop tokens
 
-    # Coba gaya call paling umum
     try:
         return model.generate(text, **kwargs)
     except TypeError:
@@ -199,7 +198,9 @@ def _generate_with_safe_kwargs(model, text: str, prompt_path: str):
             return model.generate(**kwargs)
         return model.generate(text)
 
-
+# =====================================================================
+# MAIN INFERENCE ENGINE (REFACTORED WITH INFERENCE_MODE)
+# =====================================================================
 def clone_voice(text: str, audio_file, audio_url: str, progress=gr.Progress(track_tqdm=False)):
     try:
         raw_text = (text or "").strip()
@@ -220,43 +221,55 @@ def clone_voice(text: str, audio_file, audio_url: str, progress=gr.Progress(trac
         if not chunks:
             raise gr.Error("Gagal memproses teks (chunk kosong).")
 
+        # FILTER SMART: Buang fragmen chunk yang tidak mengandung karakter alfanumerik murni
+        chunks = [ch for ch in chunks if re.search(r'[a-zA-Z0-9]', ch)]
+
         if len(chunks) > MAX_CHUNKS:
             raise gr.Error(
-                f"Teks terlalu panjang ({len(chunks)} chunk). "
-                f"Maksimal {MAX_CHUNKS} chunk per request. "
-                "Silakan pecah teks jadi beberapa bagian."
+                f"Teks terlalu padat ({len(chunks)} chunk). "
+                f"Maksimal {MAX_CHUNKS} chunk per request."
             )
 
         model = get_model()
         sr = getattr(model, "sr", 24000)
 
+        # Kunci manual seed di luar perulangan
         torch.manual_seed(42)
 
         wav_parts = []
         pause = torch.zeros(1, int(sr * PAUSE_SECONDS))
 
         total = len(chunks)
-        with torch.no_grad():
+        
+        # MANFAATKAN INFERENCE MODE (Jauh lebih hemat RAM & cepat dibanding no_grad)
+        with torch.inference_mode():
             for i, ch in enumerate(chunks, start=1):
-                progress((i - 1) / total, desc=f"Processing chunk {i}/{total}...")
+                progress((i - 1) / total, desc=f"Memproses chunk biner {i}/{total}...")
                 ch = _prepare_text_exact(ch)
 
                 wav = _generate_with_safe_kwargs(model, ch, prompt_path)
                 if wav.dim() == 1:
                     wav = wav.unsqueeze(0)
 
-                wav_parts.append(wav.cpu())
+                # Segera pindahkan hasil tensor ke penampung cpu bersih untuk memotong graph memori
+                wav_parts.append(wav.detach().cpu().clone())
                 wav_parts.append(pause)
 
-        # buang pause terakhir
+        # Buang elemen pause buatan di bagian akhir audio tracker
         if wav_parts:
             wav_parts = wav_parts[:-1]
 
+        progress(0.95, desc="Menggabungkan seluruh komponen bytes audio...")
         full_wav = torch.cat(wav_parts, dim=1)
 
         out_path = tempfile.NamedTemporaryFile(delete=False, suffix=".wav").name
         ta.save(out_path, full_wav, sr)
 
+        # PEMBERSIHAN AGRESIF: Paksa Garbage Collector Python membuang sisa array tensor
+        del wav_parts
+        del full_wav
+        gc.collect()
+
         progress(1.0, desc="Selesai ✅")
         return out_path
 
@@ -265,38 +278,38 @@ def clone_voice(text: str, audio_file, audio_url: str, progress=gr.Progress(trac
         print(traceback.format_exc())
         raise gr.Error(f"Gagal generate audio: {e}")
 
-
-with gr.Blocks(title="Chatterbox Indonesian Voice Cloning (CPU)") as demo:
-    gr.Markdown("## Chatterbox-TTS Indonesian (CPU)")
+# =====================================================================
+# GRADIO INTERFACE DESIGN
+# =====================================================================
+with gr.Blocks(title="Chatterbox Indonesian Voice Cloning (CPU Optimized)") as demo:
+    gr.Markdown("## EduScanner AI Voice Engine - Chatterbox CPU Optimized")
     gr.Markdown(
         f"""
-Masukkan teks + upload WAV (atau URL WAV).
-
-**Batas anti-ngaret saat ini:**
-- Maks total teks: **{MAX_TOTAL_CHARS}** karakter
-- Maks per chunk: **{MAX_CHARS_PER_CHUNK}** karakter
-- Maks chunk: **{MAX_CHUNKS}**
+**Spesifikasi Keamanan Sumber Daya (Hugging Face Shared CPU Tier):**
+- Batas Maksimal Karakter Masuk: **{MAX_TOTAL_CHARS}** karakter.
+- Pembagian Kapasitas per Chunk: **{MAX_CHARS_PER_CHUNK}** karakter.
+- Manajemen Batas Atas Antrean Fragmentasi: **{MAX_CHUNKS}** chunk.
 """
     )
 
     text_in = gr.Textbox(
-        label="Text Prompt",
+        label="Teks Akademik / Rangkuman (Maks 2400 Karakter)",
         lines=8,
-        placeholder="Contoh: Materi ini membahas data mining..."
+        placeholder="Masukkan teks rangkuman materi perkuliahan di sini untuk diubah menjadi suara kloning..."
     )
 
     wav_in = gr.Audio(
-        label="Upload WAV Prompt",
+        label="Sampel Suara Target (Upload WAV)",
         type="filepath"
     )
 
     url_in = gr.Textbox(
-        label="Audio URL WAV (opsional)",
-        placeholder="https://example.com/input.wav"
+        label="Atau Gunakan URL Sampel Audio WAV (Opsional)",
+        placeholder="https://domain-kamu.com/assets/sample_suara.wav"
     )
 
-    btn = gr.Button("Generate")
-    out_audio = gr.Audio(label="Hasil Audio", type="filepath")
+    btn = gr.Button("Mulai Sintesis Suara", variant="primary")
+    out_audio = gr.Audio(label="Hasil Audio Hasil Kloning (WAV)", type="filepath")
 
     btn.click(
         fn=clone_voice,
@@ -307,5 +320,6 @@ Masukkan teks + upload WAV (atau URL WAV).
 
 if __name__ == "__main__":
     port = int(os.getenv("PORT", "7860"))
+    # default_concurrency_limit=1 memaksa antrean global agar antarmuka tidak crash kelebihan beban
     demo.queue(default_concurrency_limit=1)
-    demo.launch(server_name="0.0.0.0", server_port=port, show_error=True)
+    demo.launch(server_name="0.0.0.0", server_port=port, show_error=True)