Upload gen_samples.py with huggingface_hub

gen_samples.py

@@ -32,11 +32,16 @@ def load_ref_audio(ref_path):
     return wav
 
 def encode_ref(codec, wav):
-    wav = wav.to(DEVICE)
-    with torch.autocast(device_type="cuda", dtype=DTYPE):
-        enc = codec.encode(wav.unsqueeze(0))
-        tokens = enc[0] if isinstance(enc, tuple) else enc
-        return tokens.cpu().numpy()
+    """Encode reference audio outside inference mode."""
+    import torch
+    # Exit inference mode temporarily for codec (it needs autograd for some ops)
+    wav_clone = wav.clone().detach().to(DEVICE).requires_grad_(False)
+    # Use torch.no_grad context instead of inference_mode
+    with torch.no_grad():
+        with torch.amp.autocast(device_type="cuda", dtype=DTYPE):
+            enc = codec.encode(wav_clone.unsqueeze(0))
+    tokens = enc[0] if isinstance(enc, tuple) else enc
+    return tokens.detach().cpu().numpy()
 
 def generate_clone(model, codec, ref_tokens, ref_text, gen_text, out_path):
     """Generate voice clone using the Conversation API correctly."""