perf: optimize model transfer with non-blocking flag

Author: MrPandir

app/model.py

@@ -130,7 +130,9 @@ def wrapper(*args, **kwargs):
 
 
 tokenizer = AutoTokenizer.from_pretrained(model_path)
-model = AutoModelForSequenceClassification.from_pretrained(model_path).to(device)
+model = AutoModelForSequenceClassification.from_pretrained(model_path).to(
+    device, non_blocking=True
+)
 
 
 @measure_time