Update app.py

app.py

@@ -7,8 +7,8 @@ from threading import Thread
 from accelerate import init_empty_weights
 
 max_memory = {
-    0: "50GiB",  # بسته به VRAM کارت گرافیکت، مثلاً اگه 8 گیگ داری، بذار 6 یا 5
-    "cpu": "32GiB",  # بسته به RAM سیستمت
+    0: "50GiB", 
+    "cpu": "32GiB",  
 }
 MODEL_LIST = ["THUDM/GLM-4-Z1-32B-0414"]
 
@@ -51,7 +51,7 @@ def stream_chat(message, history: list, temperature: float, max_length: int, top
         low_cpu_mem_usage=True,
         trust_remote_code=True,
         quantization_config=quantization_config,
-        device_map="auto"و
+        device_map="auto",
         max_memory=max_memory,
     )