Update GPT_SoVITS/app.py

GPT_SoVITS/app.py

@@ -428,16 +428,20 @@ def vc_main(wav_path, text, language, prompt_wav, noise_scale=0.5):
     phones, word2ph, norm_text = get_cleaned_text_final(text, language)
 
     spec = get_spepc(hps, prompt_wav) 
-    codes = get_code_from_wav(wav_path)[None, None]  # 必须是 3D, [n_q, B, T]
+    spec = spec.to(device)
+    codes = get_code_from_wav(wav_path)[None, None].to(device)  # 必须是 3D, [n_q, B, T]
     ge = vq_model.ref_enc(spec)  # [B, D, T/1] 
     quantized = vq_model.quantizer.decode(codes)  # [B, D, T]
     if hps.model.semantic_frame_rate == "25hz":
         quantized = F.interpolate(
             quantized, size=int(quantized.shape[-1] * 2), mode="nearest"
         )
+    lengths_tensor = torch.LongTensor([quantized.shape[-1]]).to(device)
+    phones_tensor = torch.LongTensor(phones)[None].to(device)
+    phones_lengths_tensor = torch.LongTensor([len(phones)]).to(device)
+
     _, m_p, logs_p, y_mask = vq_model.enc_p(
-        quantized, torch.LongTensor([quantized.shape[-1]]), 
-        torch.LongTensor(phones)[None], torch.LongTensor([len(phones)]), ge
+        quantized, lengths_tensor, phones_tensor, phones_lengths_tensor, ge
     )
     z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
     z = vq_model.flow(z_p, y_mask, g=ge, reverse=True)