Update app.py

app.py

@@ -5,8 +5,8 @@ import os
 MODEL_ID = "Muhammadidrees/MedicalInsights"
 
 # Create offload directory inside Space
-offload_dir = "./offload"
-os.makedirs(offload_dir, exist_ok=True)
+#offload_dir = "./offload"
+#os.makedirs(offload_dir, exist_ok=True)
 
 tokenizer = AutoTokenizer.from_pretrained(
     MODEL_ID,
@@ -15,8 +15,8 @@ tokenizer = AutoTokenizer.from_pretrained(
 
 model = AutoModelForCausalLM.from_pretrained(
     MODEL_ID,
-    device_map="auto",          # still balances CPU/GPU
-    offload_folder=offload_dir, # where to dump layers too big for RAM
+    device_map={"": "cpu"},
+    low_cpu_mem_usage=True,
     trust_remote_code=True
 )