Update app.py

app.py

@@ -1,4 +1,3 @@
-import spaces
 import gradio as gr
 import torch
 from TTS.api import TTS
@@ -12,6 +11,9 @@ from TTS.tts.models.xtts import Xtts
 # Aceptar los términos de COQUI
 os.environ["COQUI_TOS_AGREED"] = "1"
 
+# Establecer precisión reducida para acelerar en CPU
+torch.set_default_dtype(torch.float16)
+
 # Definir el dispositivo como CPU
 device = "cpu"
 
@@ -20,28 +22,19 @@ model_path = hf_hub_download(repo_id="RedSparkie/danielmula", filename="model.pt
 config_path = hf_hub_download(repo_id="RedSparkie/danielmula", filename="config.json")
 vocab_path = hf_hub_download(repo_id="RedSparkie/danielmula", filename="vocab.json")
 
-# Función para limpiar la caché de GPU (no necesaria para CPU, pero la mantengo por si en el futuro usas GPU)
-def clear_gpu_cache():
-    if torch.cuda.is_available():
-        torch.cuda.empty_cache()
-
 # Cargar el modelo XTTS
 XTTS_MODEL = None
 def load_model(xtts_checkpoint, xtts_config, xtts_vocab):
     global XTTS_MODEL
-    clear_gpu_cache()
-    if not xtts_checkpoint or not xtts_config or not xtts_vocab:
-        return "You need to run the previous steps or manually set the `XTTS checkpoint path`, `XTTS config path`, and `XTTS vocab path` fields !!"
     config = XttsConfig()
     config.load_json(xtts_config)
+    
+    # Inicializar el modelo
     XTTS_MODEL = Xtts.init_from_config(config)
     print("Loading XTTS model!")
-    XTTS_MODEL.load_checkpoint(config, checkpoint_path=xtts_checkpoint, vocab_path=xtts_vocab, use_deepspeed=False)
     
-    # No mover a GPU ya que usamos CPU
-    # if torch.cuda.is_available():
-    #     XTTS_MODEL.cuda()
-
+    # Cargar el checkpoint del modelo con `weights_only=True` para evitar advertencias de seguridad
+    XTTS_MODEL.load_checkpoint(config, checkpoint_path=xtts_checkpoint, vocab_path=xtts_vocab, use_deepspeed=False, weights_only=True)
     print("Model Loaded!")
 
 # Función para ejecutar TTS
@@ -49,24 +42,28 @@ def run_tts(lang, tts_text, speaker_audio_file):
     if XTTS_MODEL is None or not speaker_audio_file:
         return "You need to run the previous step to load the model !!", None, None
 
-    gpt_cond_latent, speaker_embedding = XTTS_MODEL.get_conditioning_latents(
-        audio_path=speaker_audio_file, 
-        gpt_cond_len=XTTS_MODEL.config.gpt_cond_len, 
-        max_ref_length=XTTS_MODEL.config.max_ref_len, 
-        sound_norm_refs=XTTS_MODEL.config.sound_norm_refs
-    )
-    out = XTTS_MODEL.inference(
-        text=tts_text,
-        language=lang,
-        gpt_cond_latent=gpt_cond_latent,
-        speaker_embedding=speaker_embedding,
-        temperature=XTTS_MODEL.config.temperature,
-        length_penalty=XTTS_MODEL.config.length_penalty,
-        repetition_penalty=XTTS_MODEL.config.repetition_penalty,
-        top_k=XTTS_MODEL.config.top_k,
-        top_p=XTTS_MODEL.config.top_p,
-    )
-
+    # Usar inference_mode para mejorar el rendimiento
+    with torch.inference_mode():
+        gpt_cond_latent, speaker_embedding = XTTS_MODEL.get_conditioning_latents(
+            audio_path=speaker_audio_file, 
+            gpt_cond_len=XTTS_MODEL.config.gpt_cond_len, 
+            max_ref_length=XTTS_MODEL.config.max_ref_len, 
+            sound_norm_refs=XTTS_MODEL.config.sound_norm_refs
+        )
+        
+        out = XTTS_MODEL.inference(
+            text=tts_text,
+            language=lang,
+            gpt_cond_latent=gpt_cond_latent,
+            speaker_embedding=speaker_embedding,
+            temperature=XTTS_MODEL.config.temperature,
+            length_penalty=XTTS_MODEL.config.length_penalty,
+            repetition_penalty=XTTS_MODEL.config.repetition_penalty,
+            top_k=XTTS_MODEL.config.top_k,
+            top_p=XTTS_MODEL.config.top_p,
+        )
+    
+    # Guardar el audio generado en un archivo temporal
     with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as fp:
         out["wav"] = torch.tensor(out["wav"]).unsqueeze(0)
         out_path = fp.name
@@ -76,7 +73,6 @@ def run_tts(lang, tts_text, speaker_audio_file):
     return out_path, speaker_audio_file
 
 # Definir la función para Gradio
-@spaces.GPU(enable_queue=True)
 def generate(text, audio):
     load_model(model_path, config_path, vocab_path)
     out_path, speaker_audio_file = run_tts(lang='es', tts_text=text, speaker_audio_file=audio)
@@ -85,7 +81,7 @@ def generate(text, audio):
 # Configurar la interfaz de Gradio
 demo = gr.Interface(
     fn=generate, 
-    inputs=[gr.Textbox(label='Frase a generar'), gr.Audio(type='filepath', label='Voz de referencia')], 
+    inputs=[gr.Textbox(label='Texto:'), gr.Audio(type='filepath', label='Voz de referencia')], 
     outputs=gr.Audio(type='filepath')
 )