Upload folder using huggingface_hub

app.py

@@ -174,31 +174,9 @@ RUNNER = OptimizedTTSRunner(
 SAMPLE_RATE = int(RUNNER.sample_rate)
 print(f"Model loaded. sample_rate={SAMPLE_RATE}", flush=True)
 
-# Prewarm the model with a short dummy generation so the first user request
-# does not pay the torch.compile cost.
-print("Prewarming model …", flush=True)
-_prewarm_start = time.perf_counter()
-try:
-    _pw = RUNNER.synthesize(
-        text="The optimized demo is ready.",
-        language="en",
-        speaker_id=31,
-        max_new_tokens=64,
-        text_temperature=0.0,
-        text_top_p=1.0,
-        text_top_k=None,
-        audio_temperature=0.8,
-        audio_top_p=0.92,
-        audio_top_k=None,
-        audio_repetition_penalty=1.05,
-        n_vq_for_inference=32,
-        style_text="The optimized demo is ready.",
-        style_emotion_id=0,
-        style_energy=0.5,
-    )
-    print(f"Prewarm complete in {time.perf_counter() - _prewarm_start:.1f}s", flush=True)
-except Exception as exc:
-    print(f"Prewarm failed ({exc!r}) — continuing anyway", flush=True)
+# NOTE: prewarm is skipped at module scope on ZeroGPU — no real GPU is
+# attached to the main process. The first @spaces.GPU call will pay the
+# compile/warmup cost.
 
 
 # ── Helpers ────────────────────────────────────────────────────────────────

portable_tts_runtime.py

@@ -192,7 +192,10 @@ class TorchDecoder4FeatureRuntime:
         self.inputs = [SimpleNamespace(name="codes"), SimpleNamespace(name="lengths")]
         del audio_tokenizer
         gc.collect()
-        torch.cuda.empty_cache()
+        try:
+            torch.cuda.empty_cache()
+        except Exception:
+            pass
 
     def get_inputs(self) -> list[Any]:
         return self.inputs
@@ -237,7 +240,13 @@ class TorchScriptVocoderRuntime:
         if not artifact.exists():
             raise FileNotFoundError(f"Missing TorchScript vocoder artifact: {artifact}")
         self.device = device
-        self.module = torch.jit.load(str(artifact), map_location=device).eval()
+        # Load on CPU first, then move to device via .to() so ZeroGPU's
+        # torch.cuda hijack intercepts the move (torch.jit.load with
+        # map_location="cuda" bypasses the hijack and fails in the main
+        # process where no real GPU is attached).
+        self.module = torch.jit.load(str(artifact), map_location="cpu").eval()
+        if device.type == "cuda":
+            self.module = self.module.to(device)
         self.use_cudagraph = bool(use_cudagraph and device.type == "cuda")
         self.bucket_frames = max(0, int(bucket_frames))
         self.graphs: dict[tuple[Any, ...], tuple[torch.cuda.CUDAGraph, torch.Tensor, torch.Tensor]] = {}
@@ -503,7 +512,12 @@ class OptimizedTTSRunner:
         if isinstance(self.vocoder_session, TorchScriptVocoderRuntime):
             prewarm_buckets = parse_int_csv(config.vocoder_prewarm_buckets)
             if prewarm_buckets:
-                self.vocoder_prewarm_result = self.vocoder_session.prewarm_buckets(prewarm_buckets)
+                try:
+                    self.vocoder_prewarm_result = self.vocoder_session.prewarm_buckets(prewarm_buckets)
+                except Exception:
+                    # Prewarm can fail in ZeroGPU main process (no real GPU
+                    # attached at startup). It's an optimization only.
+                    self.vocoder_prewarm_result = {"requested": prewarm_buckets, "warmed": [], "elapsed_sec": 0.0, "skipped": True}
         tts_load_kwargs: dict[str, Any] = {
             "trust_remote_code": True,
             "torch_dtype": self.dtype,