Update app.py

app.py

@@ -158,12 +158,12 @@ has_cuda = torch.cuda.is_available()
 
 if has_cuda:
     model_glm = (
-        AutoModel.from_pretrained(model_name, trust_remote_code=True).cuda().half()
+        AutoModel.from_pretrained(model_name, trust_remote_code=True).half().quantize(8).cuda()
     )  # 3.92G
 else:
     model_glm = AutoModel.from_pretrained(
         model_name, trust_remote_code=True
-    ).float() .half().float()
+    ).float().half().float()
 
 model_glm = model_glm.eval()