.

train_sft_ds.py

@@ -361,7 +361,6 @@ def main():
         logging_dir=logging_dir,
         do_train=True,
         do_eval=(eval_dataset is not None),
-        # evaluation_strategy / eval_strategy 通过 **ta_kwargs 传入
         eval_steps=max(50, args.save_steps // 5) 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,
@@ -376,15 +375,17 @@ def main():
         save_total_limit=2,
         deepspeed=args.deepspeed,
         dataloader_drop_last=True,
-        dataloader_num_workers=2,
+        dataloader_num_workers=0,
+        dataloader_prefetch_factor=1,
         report_to=([] if args.report_to == "none" else [args.report_to]),
         bf16=args.bf16,
         fp16=(not args.bf16),
         gradient_checkpointing=args.gradient_checkpointing,
-        remove_unused_columns=False,   # 需要保留我们的字段
+        remove_unused_columns=False,
         torch_compile=False,
         save_on_each_node=False,
-        **ta_kwargs,  # ← 兼容参数
+        logging_first_step=True,
+        **ta_kwargs,
     )
 
     trainer = Trainer(