Update tts_utils.py

tts_utils.py

@@ -3,14 +3,25 @@ import torch
 from parler_tts import ParlerTTSForConditionalGeneration
 from transformers import AutoTokenizer
 
+# Updated load_model function in tts_utils.py
 def load_model():
     model = ParlerTTSForConditionalGeneration.from_pretrained(
         "ai4bharat/indic-parler-tts",
-        torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
+        torch_dtype=torch.float32  # Force CPU-compatible dtype
     )
+    
+    # Apply dynamic quantization to Linear layers
+    quantized_model = torch.ao.quantization.quantize_dynamic(
+        model,
+        {torch.nn.Linear},  # Target layer type
+        dtype=torch.qint8
+    )
+    
     tokenizer = AutoTokenizer.from_pretrained("ai4bharat/indic-parler-tts")
     description_tokenizer = AutoTokenizer.from_pretrained("ai4bharat/indic-parler-tts")
-    return model, tokenizer, description_tokenizer
+    
+    return quantized_model, tokenizer, description_tokenizer
+
 
 def generate_speech(text, voice_prompt, model, tokenizer, description_tokenizer):
     device = "cuda" if torch.cuda.is_available() else "cpu"