cpu

inference.py

@@ -29,9 +29,10 @@ def load_model(
 
     model = AutoModelForCausalLM.from_pretrained(
         pretrained_model_name_or_path=model_name,
+        device_map="cpu",
         torch_dtype=torch.float32, # Use float32 for CPU
         low_cpu_mem_usage=True  # Helps with memory efficiency
-    ).to(device)
+    )
 
     model.eval() # Set model to evaluation mode