.

train_mm_zero3_lora.sh

@@ -3,6 +3,7 @@ FORCE_COLOR=1 deepspeed --hostfile hostfile \
   train_sft_lora.py \
     --model_name_or_path /home/test/Qwen3-32B \
     --data_glob "/home/test/datasets/my_corpus/train*.jsonl" \
+    --eval_data_glob "/home/test/jd_train/datasets/test/*.jsonl" \
     --output_dir /home/test/checkpoints/q3-32b-lora \
     --seq_len 512 \
     --bf16 \
@@ -12,9 +13,11 @@ FORCE_COLOR=1 deepspeed --hostfile hostfile \
     --warmup_ratio 0.03 \
     --lora_r 32 --lora_alpha 64 --lora_dropout 0.05 \
     --lora_exclude lm_head \
-    --max_steps 3000 \
+    --max_steps 300 \
     --log_interval 10 \
-    --eval_steps 200 \
+    --eval_steps 50 \
+    --save_steps 50 \
+    --save_total_limit 4 \
     --gradient_checkpointing \
     --deepspeed /home/test/jd_train/ds_config_zero3_lora.json \
     --report_to wandb --wandb_project ds-qwen3-lora

train_sft_lora.py

@@ -29,6 +29,7 @@ from transformers import (
 from transformers.trainer_callback import TrainerCallback
 from transformers.trainer_utils import get_last_checkpoint
 from torch.optim import AdamW as TorchAdamW
+from transformers import EarlyStoppingCallback
 
 # ==== make sure CLI ninja/nvcc are reachable even in non-interactive ssh ====
 import site, shutil
@@ -368,6 +369,7 @@ def parse_args():
     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)
+    ap.add_argument("--save_total_limit", type=int, default=2)
 
     # ===== LoRA 相关 =====
     ap.add_argument("--lora_r", type=int, default=16)
@@ -808,7 +810,7 @@ def main():
         lr_scheduler_type="cosine",
         logging_steps=args.log_interval,
         save_steps=args.save_steps,
-        save_total_limit=2,
+        save_total_limit=args.save_total_limit,
         deepspeed=(args.deepspeed if use_ds else None),
         dataloader_drop_last=False,
         dataloader_num_workers=0,
@@ -827,6 +829,12 @@ def main():
         ta_kwargs2["torch_compile"] = False
     ta_kwargs2.update({"bf16": (dtype==torch.bfloat16), "fp16": (dtype==torch.float16)})
 
+    ta_kwargs2.update(dict(
+        load_best_model_at_end=True,
+        metric_for_best_model="eval_loss",
+        greater_is_better=False,
+    ))
+
     training_args = TrainingArguments(**ta_kwargs2)
 
     trainer_kwargs = {}
@@ -893,6 +901,8 @@ def main():
             print(f"[resume] {resume_flag} not found locally -> disable resume.", flush=True)
             resume_flag = None
 
+    trainer.add_callback(EarlyStoppingCallback(early_stopping_patience=3, early_stopping_threshold=1e-3))
+
     print_once(f"[resume] final = {resume_flag if resume_flag else 'None (fresh start)'}")
     print_once("***** Starting LoRA training *****")