Update app.py

app.py

@@ -51,7 +51,12 @@ def quantize_llama(model_path, quantized_output_path, quant_type):
     quantize_path = f"{model_dir_path}/build/bin/llama-quantize"
     subprocess.run(["chmod", "+x", quantize_path], check=True)
     
-    cmd = [quantize_path, model_path, quantized_output_path, quant_type]
+    cmd = [
+    f"{model_dir_path}/build/bin/llama-quantize", 
+    model_path, 
+    quantized_output_path,
+    quant_type
+    ]
     
     process = subprocess.run(cmd, text=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE)