.

mm-zero3.sh

@@ -41,5 +41,6 @@ deepspeed --hostfile hostfile \
     --bf16 \
     --deepspeed /home/test/jd_train/ds_config_zero3.json \
     --report_to wandb \
+    --eval_steps 10 \
     --eval_data_glob "/home/test/datasets/my_corpus/test.jsonl"
 

train_sft_ds.py

@@ -887,6 +887,8 @@ def parse_args():
                     help="for deepspeed/torchrun launcher; ignored by user code")
     ap.add_argument("--per_device_eval_batch_size", type=int, default=1)
     ap.add_argument("--deepspeed", type=str, default=None)
+    ap.add_argument("--eval_steps", type=int, default=10,
+                    help="Evaluate every N optimizer steps when eval_dataset is provided")    
     
     return ap.parse_args()
 
@@ -1240,7 +1242,9 @@ def main():
         logging_dir=logging_dir,
         do_train=True,
         do_eval=(eval_dataset is not None),
-        eval_steps=max(50, args.save_steps // 5) if eval_dataset is not None else None,
+        # eval_steps=max(50, args.save_steps // 5) if eval_dataset is not None else None,
+        # 用用户指定的 eval_steps；没有 eval 集就 None
+        eval_steps=(args.eval_steps if eval_dataset is not None else None),        
         per_device_train_batch_size=args.per_device_train_batch_size,
         gradient_accumulation_steps=args.gradient_accumulation_steps,
         learning_rate=args.learning_rate,