Update train.py

train.py

@@ -121,6 +121,7 @@ class DistillationTrainer(Trainer):
         return loss
 
 # 训练参数
+# ✅ 移除 `use_cache` 选项
 training_args = TrainingArguments(
     output_dir="/tmp/distilled_model",
     num_train_epochs=3,
@@ -132,10 +133,12 @@ training_args = TrainingArguments(
     save_strategy="epoch",
     remove_unused_columns=False,
     gradient_checkpointing=True,
-    use_cache=False,  # ✅ **修正 `use_cache=True` 的冲突**
     fp16=torch.cuda.is_available()
 )
 
+# ✅ **手动禁用 `use_cache`**
+student.config.use_cache = False  # 🔥 这样就不会影响 `TrainingArguments`，但依然禁用了 `use_cache`
+
 # 初始化 Trainer
 trainer = DistillationTrainer(
     teacher=teacher,