diff --git a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py
index 441b4f266..c668e7f64 100755
--- a/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py
+++ b/egs/speech_llm/ASR_LLM/zipformer_llm_zh/train.py
@@ -685,6 +685,30 @@ def train_one_epoch(
                 valid_info.write_summary(
                     tb_writer, "train/valid_", params.batch_idx_train
                 )
+            if batch_idx != 0:
+                model.save_checkpoint(
+                    save_dir=params.exp_dir,
+                    tag=f"epoch-{params.cur_epoch}-checkpoint-{batch_idx}",
+                    client_state={},
+                    exclude_frozen_parameters=True,
+                )
+
+                if rank == 0:
+                    convert_zero_checkpoint_to_fp32_state_dict(
+                        params.exp_dir,
+                        f"{params.exp_dir}/epoch-{params.cur_epoch}-checkpoint-{batch_idx}.pt",
+                        tag=f"epoch-{params.cur_epoch}-checkpoint-{batch_idx}",
+                        exclude_frozen_parameters=True,
+                    )
+                    # save sampler state dict into checkpoint
+                    sampler_state_dict = train_dl.sampler.state_dict()
+                    torch.save(
+                        sampler_state_dict,
+                        f"{params.exp_dir}/epoch-{params.cur_epoch}-checkpoint-{batch_idx}-sampler.pt",
+                    )
+                    os.system(
+                        f"rm -rf {params.exp_dir}/epoch-{params.cur_epoch}-checkpoint-{batch_idx}"
+                    )
 
         shave_rate = params.shave_rate
         while True: