Upload app.py with huggingface_hub

app.py

@@ -18,9 +18,7 @@ from irodori_tts.inference_runtime import (
 # Configuration
 # ---------------------------------------------------------------------------
 
-MODEL_REPO = os.environ.get(
-    "MODEL_REPO", "Aratako/Irodori-TTS-v4.1-Small-Quantized/int8-weight-only"
-)
+MODEL_REPO = os.environ.get("MODEL_REPO", "Aratako/Irodori-TTS-v4.1-Small")
 CODEC_REPO = "Aratako/Semantic-DACVAE-Japanese-32dim"
 MAX_GRADIO_CANDIDATES = int(os.environ.get("MAX_GRADIO_CANDIDATES", "8"))
 GRADIO_AUDIO_COLS_PER_ROW = 4
@@ -129,6 +127,8 @@ def _load_models():
 # Load models at startup (ZeroGPU intercepts .to("cuda"))
 _load_models()
 
+_quantized = False
+
 
 # ---------------------------------------------------------------------------
 # GPU-decorated Inference
@@ -164,6 +164,9 @@ def run_inference_gpu(
 ) -> tuple[list[tuple[int, np.ndarray]], str]:
     """Generate Japanese speech from text using Irodori-TTS-v4.1-Small (INT8 quantized).
 
+    The base bf16 model is loaded at startup; INT8 weight-only quantization is
+    applied on the first GPU call to reduce VRAM usage while preserving quality.
+
     Args:
         text: Japanese text to synthesize.
         caption: Optional style/emotion prompt.
@@ -196,6 +199,16 @@ def run_inference_gpu(
     if _runtime is None:
         _load_models()
 
+    global _quantized
+    if not _quantized:
+        from irodori_tts.quantization import quantize_model
+
+        quantized_count = quantize_model(
+            _runtime.model, quantization_type="int8_weight_only", profile="core"
+        )
+        print(f"[Info] Quantized {quantized_count} layers to INT8 weight-only.", flush=True)
+        _quantized = True
+
     log_buffer = io.StringIO()
 
     def stdout_log(msg: str) -> None: