perf: Optimize CPU usage by setting PyTorch threads to 1 and enabling low memory model loading, and remove an unnecessary return.

app.py

@@ -8,6 +8,8 @@ import torch.nn as nn
 import psutil
 import gc
 
+torch.set_num_threads(1)
+
 # === CUSTOM MODEL CLASSES ===
 class WhisperEncoderOnlyConfig(WhisperConfig):
     model_type = "whisper_encoder_classifier"
@@ -78,7 +80,7 @@ MODEL_REPO = "tachiwin/language_classification_enconly_model_2"
 
 print("Loading model on CPU...")
 processor = WhisperProcessor.from_pretrained(MODEL_REPO)
-model = WhisperEncoderOnlyForClassification.from_pretrained(MODEL_REPO)
+model = WhisperEncoderOnlyForClassification.from_pretrained(MODEL_REPO, low_cpu_mem_usage=True)
 model.eval()
 
 print("Model loaded successfully!")
@@ -131,7 +133,7 @@ def predict_language(audio):
         return_tensors="pt",
     )
     print("[LOG] DID EXTRACT")
-    return None
+    
     # Delete raw audio array immediately as it's now in 'inputs'
     del audio_array
     gc.collect()