Update app.py

app.py

@@ -53,14 +53,39 @@ projection = torch.nn.Linear(clip_embed, phi_embed).to(device)
 resblock = SimpleResBlock(phi_embed).to(device)
 phi_model = AutoModelForCausalLM.from_pretrained(phi_model_name,trust_remote_code=True).to(device)
 # Load the model with the appropriate compute_type
+# Load the audio model with appropriate compute_type
 audio_model_size = "tiny"
+compute_type = "float32"  # Ensure using a compatible compute type
 try:
-    audio_model = whisperx.load_model(audio_model_size, device, compute_type=compute_type)
+    audio_model = whisperx.load_model(
+        audio_model_size,
+        device,
+        compute_type=compute_type,
+        # Provide necessary parameters based on your version of whisperx
+        # Check documentation for required parameters
+        max_new_tokens=100,  # Example values, adjust as needed
+        clip_timestamps=True,
+        hallucination_silence_threshold=0.5,
+        hotwords=None  # Add specific hotwords if needed
+    )
     print(f"Model loaded successfully with compute_type: {compute_type}")
 except ValueError as e:
     print(f"Error loading model: {e}")
-    print("Falling back to int8 compute type")
-    audio_model = whisperx.load_model(audio_model_size, device, compute_type="int8")
+    # Optionally, try loading with int8 if necessary
+    try:
+        audio_model = whisperx.load_model(
+            audio_model_size,
+            device,
+            compute_type="int8",
+            max_new_tokens=100,
+            clip_timestamps=True,
+            hallucination_silence_threshold=0.5,
+            hotwords=None
+        )
+        print("Fell back to int8 compute type successfully.")
+    except Exception as e:
+        print(f"Failed to load model with int8: {e}")
+
 
 # load weights
 model_to_merge = PeftModel.from_pretrained(phi_model,'./model_chkpt/lora_adaptor')