Update app.py

app.py

@@ -6,6 +6,13 @@ from transformers import AutoTokenizer
 import soundfile as sf
 import tempfile
 
+# Load model and tokenizers at startup (on CPU initially)
+print("Loading model and tokenizers...")
+model = ParlerTTSForConditionalGeneration.from_pretrained("ai4bharat/indic-parler-tts").to("cpu")
+tokenizer = AutoTokenizer.from_pretrained("ai4bharat/indic-parler-tts")
+description_tokenizer = AutoTokenizer.from_pretrained(model.config.text_encoder._name_or_path)
+print("Model and tokenizers loaded.")
+
 # Supported languages and default settings
 languages = {
     "Urdu": "A female speaker delivers a clear and expressive speech in Urdu.",
@@ -33,10 +40,10 @@ def generate_description(language, gender, emotion, noise, reverb, expressivity,
 # Generate audio function with GPU allocation
 @spaces.GPU  # Allocate GPU for the duration of this function
 def generate_audio(text, description):
-    # Load model and tokenizer
-    model = ParlerTTSForConditionalGeneration.from_pretrained("ai4bharat/indic-parler-tts").to("cuda")
-    tokenizer = AutoTokenizer.from_pretrained("ai4bharat/indic-parler-tts")
-    description_tokenizer = AutoTokenizer.from_pretrained(model.config.text_encoder._name_or_path)
+    global model  # Access the preloaded model
+
+    # Move model to GPU
+    model.to("cuda")
 
     # Prepare model inputs
     input_ids = description_tokenizer(description, return_tensors="pt").input_ids.to("cuda")
@@ -51,6 +58,9 @@ def generate_audio(text, description):
         sf.write(f.name, audio_arr, model.config.sampling_rate)
         audio_path = f.name
 
+    # Move model back to CPU to free GPU memory
+    model.to("cpu")
+
     return audio_path
 
 # Gradio Interface