diff --git a/fastdeploy/model_executor/models/glm4_moe.py b/fastdeploy/model_executor/models/glm4_moe.py
index 01b38bedd..40a492126 100644
--- a/fastdeploy/model_executor/models/glm4_moe.py
+++ b/fastdeploy/model_executor/models/glm4_moe.py
@@ -17,9 +17,12 @@
 from __future__ import annotations
 
 import re
+from functools import partial
 
 import paddle
 from paddle import nn
+from paddleformers.transformers import PretrainedModel
+from paddleformers.utils.log import logger
 
 from fastdeploy.config import FDConfig
 from fastdeploy.distributed.communication import tensor_model_parallel_all_reduce
@@ -504,3 +507,86 @@ class Glm4MoeForCausalLM(ModelForCasualLM):
     def clear_grpah_opt_backend(self):
         """Clear graph optimization backend, the captured cuda graph will be cleaned"""
         self.model.clear_grpah_opt_backend(fd_config=self.fd_config)
+
+
+class Glm4MoePretrainedModel(PretrainedModel):
+    """
+    Glm4MoePretrainedModel
+    """
+
+    config_class = FDConfig
+
+    def _init_weight(self, layer):
+        """
+        _init_weight
+        """
+        return None
+
+    @classmethod
+    def arch_name(self):
+        return "Glm4MoeForCausalLM"
+
+    @classmethod
+    def _get_tensor_parallel_mappings(cls, config, is_split=True):
+
+        logger.info("Glm4Moe inference model _get_tensor_parallel_mappings")
+
+        from fastdeploy.model_executor.models.tp_utils import split_or_merge_func_v1
+
+        fn = split_or_merge_func_v1(
+            is_split=is_split,
+            tensor_parallel_degree=config.tensor_parallel_degree,
+            tensor_parallel_rank=config.tensor_parallel_rank,
+            num_attention_heads=config.num_attention_heads,
+            num_key_value_heads=config.num_key_value_heads,
+            head_dim=config.head_dim,
+        )
+
+        def get_tensor_parallel_split_mappings(num_layers):
+            final_actions = {}
+
+            base_actions = {
+                "lm_head.weight": partial(fn, is_column=True),
+                "embed_tokens.weight": partial(fn, is_column=False),
+                "layers.0.self_attn.o_proj.weight": partial(fn, is_column=False),
+            }
+
+            # Self Attention Layer which are need TP.
+            base_actions["layers.0.self_attn.q_proj.weight"] = partial(fn, is_column=True)
+            base_actions["layers.0.self_attn.k_proj.weight"] = partial(fn, is_column=True)
+            base_actions["layers.0.self_attn.v_proj.weight"] = partial(fn, is_column=True)
+            base_actions["layers.0.self_attn.q_proj.bias"] = partial(fn, is_column=True)
+            base_actions["layers.0.self_attn.k_proj.bias"] = partial(fn, is_column=True)
+            base_actions["layers.0.self_attn.v_proj.bias"] = partial(fn, is_column=True)
+
+            # MLP Layer
+            base_actions["layers.0.mlp.gate_proj.weight"] = partial(fn, is_column=True)
+            base_actions["layers.0.mlp.up_proj.weight"] = partial(fn, is_column=True)
+            base_actions["layers.0.mlp.down_proj.weight"] = partial(fn, is_column=False)
+
+            # Moe Layer
+            for expert_idx in range(config.n_routed_experts):
+                base_actions[f"layers.0.mlp.experts.{expert_idx}.up_proj.weight"] = partial(fn, is_column=True)
+                base_actions[f"layers.0.mlp.experts.{expert_idx}.gate_proj.weight"] = partial(fn, is_column=True)
+                base_actions[f"layers.0.mlp.experts.{expert_idx}.down_proj.weight"] = partial(fn, is_column=False)
+
+            # Shared Expert Layer
+            base_actions["layers.0.mlp.shared_experts.up_proj.weight"] = partial(fn, is_column=True)
+            base_actions["layers.0.mlp.shared_experts.gate_proj.weight"] = partial(fn, is_column=True)
+            base_actions["layers.0.mlp.shared_experts.down_proj.weight"] = partial(fn, is_column=False)
+
+            # MTP parts
+            base_actions["layers.46.embed_tokens.weight"] = partial(fn, is_column=False)
+            base_actions["layers.46.eh_proj.weight"] = partial(fn, is_column=True)
+            base_actions["layers.46.shared_head.head.weight"] = partial(fn, is_column=True)
+
+            for key, action in base_actions.items():
+                if "layers.0." in key:
+                    for i in range(num_layers):
+                        final_actions[key.replace("layers.0.", f"layers.{i}.")] = action
+                final_actions[key] = action
+
+            return final_actions
+
+        mappings = get_tensor_parallel_split_mappings(config.num_hidden_layers)
+        return mappings
diff --git a/fastdeploy/rl/rollout_model.py b/fastdeploy/rl/rollout_model.py
index 3282e4548..4e52f29e8 100644
--- a/fastdeploy/rl/rollout_model.py
+++ b/fastdeploy/rl/rollout_model.py
@@ -14,6 +14,7 @@
 # limitations under the License.
 """
 
+import copy
 from typing import Dict
 
 import paddle
@@ -28,6 +29,10 @@ from fastdeploy.model_executor.models.ernie4_5_vl.ernie4_5_vl_moe import (
     Ernie4_5_VLMoeForConditionalGeneration,
     Ernie4_5_VLPretrainedModel,
 )
+from fastdeploy.model_executor.models.glm4_moe import (
+    Glm4MoeForCausalLM,
+    Glm4MoePretrainedModel,
+)
 from fastdeploy.model_executor.models.model_base import ModelRegistry
 from fastdeploy.model_executor.models.qwen2 import (
     Qwen2ForCausalLM,
@@ -529,3 +534,83 @@ class Qwen2_5_VLForConditionalGenerationRL(Qwen2_5_VLForConditionalGeneration, B
         self._complete_missing_mappings()
 
         return self.infer_to_train_mapping
+
+
+class Glm4MoeForCausalLMRL(Glm4MoeForCausalLM, BaseRLModel):
+    """
+    Glm4MoeForCausalLMRL
+    """
+
+    _get_tensor_parallel_mappings = Glm4MoePretrainedModel._get_tensor_parallel_mappings
+
+    def __init__(self, fd_config: FDConfig):
+        """
+        Args:
+            fd_config (FDConfig): Configurations for the LLM model.
+        """
+        super(Glm4MoeForCausalLMRL, self).__init__(fd_config)
+
+    @classmethod
+    def name(self) -> str:
+        """name"""
+        return "Glm4MoeForCausalLMRL"
+
+    def get_name_mappings_to_training(self, trainer_degree=None) -> Dict[str, str]:
+        """Generate mapping between inference and training parameter for RL(donot delete!)."""
+        if self._mappings_built:
+            return self.infer_to_train_mapping
+
+        self.infer_to_train_mapping = {}
+        self._mappings_built = True
+        # Prepare placeholders
+        place_holders = ["weight"]
+
+        # Initialize mapping dictionary
+        self._update_base_mappings("model")
+
+        base_name = "model.layers"
+
+        # Helper function to add layer mappings
+        def _add_layer_mappings(layer_idx: int):
+            # MoE specific mappings
+            self.infer_to_train_mapping[f"{base_name}.{layer_idx}.mlp.gate.weight"] = (
+                f"{base_name}.{layer_idx}.mlp.gate.weight"
+            )
+
+            self.infer_to_train_mapping[f"{base_name}.{layer_idx}.mlp.gate.e_score_correction_bias"] = (
+                f"{base_name}.{layer_idx}.mlp.gate.e_score_correction_bias"
+            )
+
+            # MoE experts mappings
+            for expert_idx in range(self.fd_config.model_config.n_routed_experts):
+                for ph in place_holders:
+                    # up_gate_proj (up_gate_proj)
+                    up_gate_proj_key = f"{base_name}.{layer_idx}.mlp.experts.up_gate_proj_weight"
+                    if up_gate_proj_key not in self.infer_to_train_mapping:
+                        self.infer_to_train_mapping[up_gate_proj_key] = []
+                    self.infer_to_train_mapping[up_gate_proj_key].append(
+                        f"{base_name}.{layer_idx}.mlp.experts.{expert_idx}.up_gate_proj.{ph}"
+                    )
+
+                    # down_proj (down_proj)
+                    down_proj_key = f"{base_name}.{layer_idx}.mlp.experts.down_proj_weight"
+                    if down_proj_key not in self.infer_to_train_mapping:
+                        self.infer_to_train_mapping[down_proj_key] = []
+                    self.infer_to_train_mapping[down_proj_key].append(
+                        f"{base_name}.{layer_idx}.mlp.experts.{expert_idx}.down_proj.{ph}"
+                    )
+
+        # Process MoE layers
+        for layer_idx in range(
+            self.fd_config.model_config.first_k_dense_replace,
+            self.fd_config.model_config.num_hidden_layers,
+        ):
+            _add_layer_mappings(layer_idx)
+
+        self._complete_missing_mappings()
+        infer_to_train_mapping_copy = copy.deepcopy(self.infer_to_train_mapping)
+        for key in infer_to_train_mapping_copy.keys():
+            if "mlp.experts.gate_correction_bias" in key:
+                self.infer_to_train_mapping.pop(key)
+
+        return self.infer_to_train_mapping
diff --git a/tests/ci_use/EB_VL_Lite/test_rollout_model.py b/tests/ci_use/EB_VL_Lite/test_rollout_model.py
index 9fbfc4821..e9bcd93cb 100644
--- a/tests/ci_use/EB_VL_Lite/test_rollout_model.py
+++ b/tests/ci_use/EB_VL_Lite/test_rollout_model.py
@@ -22,8 +22,9 @@ def test_rollout_model_with_distributed_launch():
     test_rollout_model
     """
     current_dir = os.path.dirname(os.path.abspath(__file__))
-
-    rollout_script = os.path.join(current_dir, "rollout_model.py")
+    utils_dir = os.path.join(os.path.dirname(current_dir), "utils")
+    rollout_script = os.path.join(utils_dir, "rollout_model.py")
+    baseline_path = os.path.join(current_dir, "baseline.txt")
 
     base_path = os.getenv("MODEL_PATH")
     if base_path:
@@ -40,6 +41,11 @@ def test_rollout_model_with_distributed_launch():
         rollout_script,
         "--model_path",
         model_path,
+        "--baseline_path",
+        baseline_path,
+        "--enable_mm",
+        "--quantization",
+        "wint8",
     ]
 
     print(f"Executing command: {' '.join(command)}")
diff --git a/tests/ci_use/GLM-45-AIR/baseline.txt b/tests/ci_use/GLM-45-AIR/baseline.txt
new file mode 100644
index 000000000..7fd21998b
--- /dev/null
+++ b/tests/ci_use/GLM-45-AIR/baseline.txt
@@ -0,0 +1,43 @@
+model.embed_tokens.embeddings.weight
+model.layers.0.self_attn.qkv_proj.bias
+model.layers.0.self_attn.qkv_proj.weight
+model.layers.0.self_attn.o_proj.weight
+model.layers.0.mlp.up_gate_proj.weight
+model.layers.0.mlp.down_proj.weight
+model.layers.0.input_layernorm.weight
+model.layers.0.post_attention_layernorm.weight
+model.layers.1.self_attn.qkv_proj.bias
+model.layers.1.self_attn.qkv_proj.weight
+model.layers.1.self_attn.o_proj.weight
+model.layers.1.mlp.gate.weight
+model.layers.1.mlp.gate.e_score_correction_bias
+model.layers.1.mlp.experts.gate_correction_bias
+model.layers.1.mlp.experts.up_gate_proj_weight
+model.layers.1.mlp.experts.down_proj_weight
+model.layers.1.mlp.shared_experts.up_gate_proj.weight
+model.layers.1.mlp.shared_experts.down_proj.weight
+model.layers.1.input_layernorm.weight
+model.layers.1.post_attention_layernorm.weight
+model.norm.weight
+lm_head.linear.weight
+model.embed_tokens.embeddings.weight:model.embed_tokens.weight
+lm_head.linear.weight:lm_head.weight
+model.layers.1.mlp.gate.weight:model.layers.1.mlp.gate.weight
+model.layers.1.mlp.gate.e_score_correction_bias:model.layers.1.mlp.gate.e_score_correction_bias
+model.layers.1.mlp.experts.up_gate_proj_weight:['model.layers.1.mlp.experts.0.up_gate_proj.weight', 'model.layers.1.mlp.experts.1.up_gate_proj.weight', 'model.layers.1.mlp.experts.2.up_gate_proj.weight', 'model.layers.1.mlp.experts.3.up_gate_proj.weight', 'model.layers.1.mlp.experts.4.up_gate_proj.weight', 'model.layers.1.mlp.experts.5.up_gate_proj.weight', 'model.layers.1.mlp.experts.6.up_gate_proj.weight', 'model.layers.1.mlp.experts.7.up_gate_proj.weight', 'model.layers.1.mlp.experts.8.up_gate_proj.weight', 'model.layers.1.mlp.experts.9.up_gate_proj.weight', 'model.layers.1.mlp.experts.10.up_gate_proj.weight', 'model.layers.1.mlp.experts.11.up_gate_proj.weight', 'model.layers.1.mlp.experts.12.up_gate_proj.weight', 'model.layers.1.mlp.experts.13.up_gate_proj.weight', 'model.layers.1.mlp.experts.14.up_gate_proj.weight', 'model.layers.1.mlp.experts.15.up_gate_proj.weight', 'model.layers.1.mlp.experts.16.up_gate_proj.weight', 'model.layers.1.mlp.experts.17.up_gate_proj.weight', 'model.layers.1.mlp.experts.18.up_gate_proj.weight', 'model.layers.1.mlp.experts.19.up_gate_proj.weight', 'model.layers.1.mlp.experts.20.up_gate_proj.weight', 'model.layers.1.mlp.experts.21.up_gate_proj.weight', 'model.layers.1.mlp.experts.22.up_gate_proj.weight', 'model.layers.1.mlp.experts.23.up_gate_proj.weight', 'model.layers.1.mlp.experts.24.up_gate_proj.weight', 'model.layers.1.mlp.experts.25.up_gate_proj.weight', 'model.layers.1.mlp.experts.26.up_gate_proj.weight', 'model.layers.1.mlp.experts.27.up_gate_proj.weight', 'model.layers.1.mlp.experts.28.up_gate_proj.weight', 'model.layers.1.mlp.experts.29.up_gate_proj.weight', 'model.layers.1.mlp.experts.30.up_gate_proj.weight', 'model.layers.1.mlp.experts.31.up_gate_proj.weight', 'model.layers.1.mlp.experts.32.up_gate_proj.weight', 'model.layers.1.mlp.experts.33.up_gate_proj.weight', 'model.layers.1.mlp.experts.34.up_gate_proj.weight', 'model.layers.1.mlp.experts.35.up_gate_proj.weight', 'model.layers.1.mlp.experts.36.up_gate_proj.weight', 'model.layers.1.mlp.experts.37.up_gate_proj.weight', 'model.layers.1.mlp.experts.38.up_gate_proj.weight', 'model.layers.1.mlp.experts.39.up_gate_proj.weight', 'model.layers.1.mlp.experts.40.up_gate_proj.weight', 'model.layers.1.mlp.experts.41.up_gate_proj.weight', 'model.layers.1.mlp.experts.42.up_gate_proj.weight', 'model.layers.1.mlp.experts.43.up_gate_proj.weight', 'model.layers.1.mlp.experts.44.up_gate_proj.weight', 'model.layers.1.mlp.experts.45.up_gate_proj.weight', 'model.layers.1.mlp.experts.46.up_gate_proj.weight', 'model.layers.1.mlp.experts.47.up_gate_proj.weight', 'model.layers.1.mlp.experts.48.up_gate_proj.weight', 'model.layers.1.mlp.experts.49.up_gate_proj.weight', 'model.layers.1.mlp.experts.50.up_gate_proj.weight', 'model.layers.1.mlp.experts.51.up_gate_proj.weight', 'model.layers.1.mlp.experts.52.up_gate_proj.weight', 'model.layers.1.mlp.experts.53.up_gate_proj.weight', 'model.layers.1.mlp.experts.54.up_gate_proj.weight', 'model.layers.1.mlp.experts.55.up_gate_proj.weight', 'model.layers.1.mlp.experts.56.up_gate_proj.weight', 'model.layers.1.mlp.experts.57.up_gate_proj.weight', 'model.layers.1.mlp.experts.58.up_gate_proj.weight', 'model.layers.1.mlp.experts.59.up_gate_proj.weight', 'model.layers.1.mlp.experts.60.up_gate_proj.weight', 'model.layers.1.mlp.experts.61.up_gate_proj.weight', 'model.layers.1.mlp.experts.62.up_gate_proj.weight', 'model.layers.1.mlp.experts.63.up_gate_proj.weight', 'model.layers.1.mlp.experts.64.up_gate_proj.weight', 'model.layers.1.mlp.experts.65.up_gate_proj.weight', 'model.layers.1.mlp.experts.66.up_gate_proj.weight', 'model.layers.1.mlp.experts.67.up_gate_proj.weight', 'model.layers.1.mlp.experts.68.up_gate_proj.weight', 'model.layers.1.mlp.experts.69.up_gate_proj.weight', 'model.layers.1.mlp.experts.70.up_gate_proj.weight', 'model.layers.1.mlp.experts.71.up_gate_proj.weight', 'model.layers.1.mlp.experts.72.up_gate_proj.weight', 'model.layers.1.mlp.experts.73.up_gate_proj.weight', 'model.layers.1.mlp.experts.74.up_gate_proj.weight', 'model.layers.1.mlp.experts.75.up_gate_proj.weight', 'model.layers.1.mlp.experts.76.up_gate_proj.weight', 'model.layers.1.mlp.experts.77.up_gate_proj.weight', 'model.layers.1.mlp.experts.78.up_gate_proj.weight', 'model.layers.1.mlp.experts.79.up_gate_proj.weight', 'model.layers.1.mlp.experts.80.up_gate_proj.weight', 'model.layers.1.mlp.experts.81.up_gate_proj.weight', 'model.layers.1.mlp.experts.82.up_gate_proj.weight', 'model.layers.1.mlp.experts.83.up_gate_proj.weight', 'model.layers.1.mlp.experts.84.up_gate_proj.weight', 'model.layers.1.mlp.experts.85.up_gate_proj.weight', 'model.layers.1.mlp.experts.86.up_gate_proj.weight', 'model.layers.1.mlp.experts.87.up_gate_proj.weight', 'model.layers.1.mlp.experts.88.up_gate_proj.weight', 'model.layers.1.mlp.experts.89.up_gate_proj.weight', 'model.layers.1.mlp.experts.90.up_gate_proj.weight', 'model.layers.1.mlp.experts.91.up_gate_proj.weight', 'model.layers.1.mlp.experts.92.up_gate_proj.weight', 'model.layers.1.mlp.experts.93.up_gate_proj.weight', 'model.layers.1.mlp.experts.94.up_gate_proj.weight', 'model.layers.1.mlp.experts.95.up_gate_proj.weight', 'model.layers.1.mlp.experts.96.up_gate_proj.weight', 'model.layers.1.mlp.experts.97.up_gate_proj.weight', 'model.layers.1.mlp.experts.98.up_gate_proj.weight', 'model.layers.1.mlp.experts.99.up_gate_proj.weight', 'model.layers.1.mlp.experts.100.up_gate_proj.weight', 'model.layers.1.mlp.experts.101.up_gate_proj.weight', 'model.layers.1.mlp.experts.102.up_gate_proj.weight', 'model.layers.1.mlp.experts.103.up_gate_proj.weight', 'model.layers.1.mlp.experts.104.up_gate_proj.weight', 'model.layers.1.mlp.experts.105.up_gate_proj.weight', 'model.layers.1.mlp.experts.106.up_gate_proj.weight', 'model.layers.1.mlp.experts.107.up_gate_proj.weight', 'model.layers.1.mlp.experts.108.up_gate_proj.weight', 'model.layers.1.mlp.experts.109.up_gate_proj.weight', 'model.layers.1.mlp.experts.110.up_gate_proj.weight', 'model.layers.1.mlp.experts.111.up_gate_proj.weight', 'model.layers.1.mlp.experts.112.up_gate_proj.weight', 'model.layers.1.mlp.experts.113.up_gate_proj.weight', 'model.layers.1.mlp.experts.114.up_gate_proj.weight', 'model.layers.1.mlp.experts.115.up_gate_proj.weight', 'model.layers.1.mlp.experts.116.up_gate_proj.weight', 'model.layers.1.mlp.experts.117.up_gate_proj.weight', 'model.layers.1.mlp.experts.118.up_gate_proj.weight', 'model.layers.1.mlp.experts.119.up_gate_proj.weight', 'model.layers.1.mlp.experts.120.up_gate_proj.weight', 'model.layers.1.mlp.experts.121.up_gate_proj.weight', 'model.layers.1.mlp.experts.122.up_gate_proj.weight', 'model.layers.1.mlp.experts.123.up_gate_proj.weight', 'model.layers.1.mlp.experts.124.up_gate_proj.weight', 'model.layers.1.mlp.experts.125.up_gate_proj.weight', 'model.layers.1.mlp.experts.126.up_gate_proj.weight', 'model.layers.1.mlp.experts.127.up_gate_proj.weight']
+model.layers.1.mlp.experts.down_proj_weight:['model.layers.1.mlp.experts.0.down_proj.weight', 'model.layers.1.mlp.experts.1.down_proj.weight', 'model.layers.1.mlp.experts.2.down_proj.weight', 'model.layers.1.mlp.experts.3.down_proj.weight', 'model.layers.1.mlp.experts.4.down_proj.weight', 'model.layers.1.mlp.experts.5.down_proj.weight', 'model.layers.1.mlp.experts.6.down_proj.weight', 'model.layers.1.mlp.experts.7.down_proj.weight', 'model.layers.1.mlp.experts.8.down_proj.weight', 'model.layers.1.mlp.experts.9.down_proj.weight', 'model.layers.1.mlp.experts.10.down_proj.weight', 'model.layers.1.mlp.experts.11.down_proj.weight', 'model.layers.1.mlp.experts.12.down_proj.weight', 'model.layers.1.mlp.experts.13.down_proj.weight', 'model.layers.1.mlp.experts.14.down_proj.weight', 'model.layers.1.mlp.experts.15.down_proj.weight', 'model.layers.1.mlp.experts.16.down_proj.weight', 'model.layers.1.mlp.experts.17.down_proj.weight', 'model.layers.1.mlp.experts.18.down_proj.weight', 'model.layers.1.mlp.experts.19.down_proj.weight', 'model.layers.1.mlp.experts.20.down_proj.weight', 'model.layers.1.mlp.experts.21.down_proj.weight', 'model.layers.1.mlp.experts.22.down_proj.weight', 'model.layers.1.mlp.experts.23.down_proj.weight', 'model.layers.1.mlp.experts.24.down_proj.weight', 'model.layers.1.mlp.experts.25.down_proj.weight', 'model.layers.1.mlp.experts.26.down_proj.weight', 'model.layers.1.mlp.experts.27.down_proj.weight', 'model.layers.1.mlp.experts.28.down_proj.weight', 'model.layers.1.mlp.experts.29.down_proj.weight', 'model.layers.1.mlp.experts.30.down_proj.weight', 'model.layers.1.mlp.experts.31.down_proj.weight', 'model.layers.1.mlp.experts.32.down_proj.weight', 'model.layers.1.mlp.experts.33.down_proj.weight', 'model.layers.1.mlp.experts.34.down_proj.weight', 'model.layers.1.mlp.experts.35.down_proj.weight', 'model.layers.1.mlp.experts.36.down_proj.weight', 'model.layers.1.mlp.experts.37.down_proj.weight', 'model.layers.1.mlp.experts.38.down_proj.weight', 'model.layers.1.mlp.experts.39.down_proj.weight', 'model.layers.1.mlp.experts.40.down_proj.weight', 'model.layers.1.mlp.experts.41.down_proj.weight', 'model.layers.1.mlp.experts.42.down_proj.weight', 'model.layers.1.mlp.experts.43.down_proj.weight', 'model.layers.1.mlp.experts.44.down_proj.weight', 'model.layers.1.mlp.experts.45.down_proj.weight', 'model.layers.1.mlp.experts.46.down_proj.weight', 'model.layers.1.mlp.experts.47.down_proj.weight', 'model.layers.1.mlp.experts.48.down_proj.weight', 'model.layers.1.mlp.experts.49.down_proj.weight', 'model.layers.1.mlp.experts.50.down_proj.weight', 'model.layers.1.mlp.experts.51.down_proj.weight', 'model.layers.1.mlp.experts.52.down_proj.weight', 'model.layers.1.mlp.experts.53.down_proj.weight', 'model.layers.1.mlp.experts.54.down_proj.weight', 'model.layers.1.mlp.experts.55.down_proj.weight', 'model.layers.1.mlp.experts.56.down_proj.weight', 'model.layers.1.mlp.experts.57.down_proj.weight', 'model.layers.1.mlp.experts.58.down_proj.weight', 'model.layers.1.mlp.experts.59.down_proj.weight', 'model.layers.1.mlp.experts.60.down_proj.weight', 'model.layers.1.mlp.experts.61.down_proj.weight', 'model.layers.1.mlp.experts.62.down_proj.weight', 'model.layers.1.mlp.experts.63.down_proj.weight', 'model.layers.1.mlp.experts.64.down_proj.weight', 'model.layers.1.mlp.experts.65.down_proj.weight', 'model.layers.1.mlp.experts.66.down_proj.weight', 'model.layers.1.mlp.experts.67.down_proj.weight', 'model.layers.1.mlp.experts.68.down_proj.weight', 'model.layers.1.mlp.experts.69.down_proj.weight', 'model.layers.1.mlp.experts.70.down_proj.weight', 'model.layers.1.mlp.experts.71.down_proj.weight', 'model.layers.1.mlp.experts.72.down_proj.weight', 'model.layers.1.mlp.experts.73.down_proj.weight', 'model.layers.1.mlp.experts.74.down_proj.weight', 'model.layers.1.mlp.experts.75.down_proj.weight', 'model.layers.1.mlp.experts.76.down_proj.weight', 'model.layers.1.mlp.experts.77.down_proj.weight', 'model.layers.1.mlp.experts.78.down_proj.weight', 'model.layers.1.mlp.experts.79.down_proj.weight', 'model.layers.1.mlp.experts.80.down_proj.weight', 'model.layers.1.mlp.experts.81.down_proj.weight', 'model.layers.1.mlp.experts.82.down_proj.weight', 'model.layers.1.mlp.experts.83.down_proj.weight', 'model.layers.1.mlp.experts.84.down_proj.weight', 'model.layers.1.mlp.experts.85.down_proj.weight', 'model.layers.1.mlp.experts.86.down_proj.weight', 'model.layers.1.mlp.experts.87.down_proj.weight', 'model.layers.1.mlp.experts.88.down_proj.weight', 'model.layers.1.mlp.experts.89.down_proj.weight', 'model.layers.1.mlp.experts.90.down_proj.weight', 'model.layers.1.mlp.experts.91.down_proj.weight', 'model.layers.1.mlp.experts.92.down_proj.weight', 'model.layers.1.mlp.experts.93.down_proj.weight', 'model.layers.1.mlp.experts.94.down_proj.weight', 'model.layers.1.mlp.experts.95.down_proj.weight', 'model.layers.1.mlp.experts.96.down_proj.weight', 'model.layers.1.mlp.experts.97.down_proj.weight', 'model.layers.1.mlp.experts.98.down_proj.weight', 'model.layers.1.mlp.experts.99.down_proj.weight', 'model.layers.1.mlp.experts.100.down_proj.weight', 'model.layers.1.mlp.experts.101.down_proj.weight', 'model.layers.1.mlp.experts.102.down_proj.weight', 'model.layers.1.mlp.experts.103.down_proj.weight', 'model.layers.1.mlp.experts.104.down_proj.weight', 'model.layers.1.mlp.experts.105.down_proj.weight', 'model.layers.1.mlp.experts.106.down_proj.weight', 'model.layers.1.mlp.experts.107.down_proj.weight', 'model.layers.1.mlp.experts.108.down_proj.weight', 'model.layers.1.mlp.experts.109.down_proj.weight', 'model.layers.1.mlp.experts.110.down_proj.weight', 'model.layers.1.mlp.experts.111.down_proj.weight', 'model.layers.1.mlp.experts.112.down_proj.weight', 'model.layers.1.mlp.experts.113.down_proj.weight', 'model.layers.1.mlp.experts.114.down_proj.weight', 'model.layers.1.mlp.experts.115.down_proj.weight', 'model.layers.1.mlp.experts.116.down_proj.weight', 'model.layers.1.mlp.experts.117.down_proj.weight', 'model.layers.1.mlp.experts.118.down_proj.weight', 'model.layers.1.mlp.experts.119.down_proj.weight', 'model.layers.1.mlp.experts.120.down_proj.weight', 'model.layers.1.mlp.experts.121.down_proj.weight', 'model.layers.1.mlp.experts.122.down_proj.weight', 'model.layers.1.mlp.experts.123.down_proj.weight', 'model.layers.1.mlp.experts.124.down_proj.weight', 'model.layers.1.mlp.experts.125.down_proj.weight', 'model.layers.1.mlp.experts.126.down_proj.weight', 'model.layers.1.mlp.experts.127.down_proj.weight']
+model.layers.0.self_attn.qkv_proj.bias:model.layers.0.self_attn.qkv_proj.bias
+model.layers.0.self_attn.qkv_proj.weight:model.layers.0.self_attn.qkv_proj.weight
+model.layers.0.self_attn.o_proj.weight:model.layers.0.self_attn.o_proj.weight
+model.layers.0.mlp.up_gate_proj.weight:model.layers.0.mlp.up_gate_proj.weight
+model.layers.0.mlp.down_proj.weight:model.layers.0.mlp.down_proj.weight
+model.layers.0.input_layernorm.weight:model.layers.0.input_layernorm.weight
+model.layers.0.post_attention_layernorm.weight:model.layers.0.post_attention_layernorm.weight
+model.layers.1.self_attn.qkv_proj.bias:model.layers.1.self_attn.qkv_proj.bias
+model.layers.1.self_attn.qkv_proj.weight:model.layers.1.self_attn.qkv_proj.weight
+model.layers.1.self_attn.o_proj.weight:model.layers.1.self_attn.o_proj.weight
+model.layers.1.mlp.shared_experts.up_gate_proj.weight:model.layers.1.mlp.shared_experts.up_gate_proj.weight
+model.layers.1.mlp.shared_experts.down_proj.weight:model.layers.1.mlp.shared_experts.down_proj.weight
+model.layers.1.input_layernorm.weight:model.layers.1.input_layernorm.weight
+model.layers.1.post_attention_layernorm.weight:model.layers.1.post_attention_layernorm.weight
+model.norm.weight:model.norm.weight
diff --git a/tests/ci_use/GLM-45-AIR/test_rollout_model.py b/tests/ci_use/GLM-45-AIR/test_rollout_model.py
new file mode 100644
index 000000000..aed74a381
--- /dev/null
+++ b/tests/ci_use/GLM-45-AIR/test_rollout_model.py
@@ -0,0 +1,66 @@
+# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import subprocess
+import sys
+
+
+def test_rollout_model_with_distributed_launch():
+    """
+    test_rollout_model
+    """
+    current_dir = os.path.dirname(os.path.abspath(__file__))
+    utils_dir = os.path.join(os.path.dirname(current_dir), "utils")
+    rollout_script = os.path.join(utils_dir, "rollout_model.py")
+    baseline_path = os.path.join(current_dir, "baseline.txt")
+
+    base_path = os.getenv("MODEL_PATH")
+    if base_path:
+        model_path = os.path.join(base_path, "GLM-4.5-Air-Fake")
+    else:
+        model_path = "./GLM-4.5-Air-Fake"
+    print(f"model_path = {model_path}")
+
+    command = [
+        sys.executable,
+        "-m",
+        "paddle.distributed.launch",
+        "--gpus",
+        "0,1",
+        rollout_script,
+        "--model_path",
+        model_path,
+        "--baseline_path",
+        baseline_path,
+    ]
+
+    print(f"Executing command: {' '.join(command)}")
+
+    process = subprocess.Popen(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+
+    try:
+        stdout, stderr = process.communicate(timeout=300)
+        return_code = process.returncode
+    except subprocess.TimeoutExpired:
+        process.kill()
+        stdout, stderr = process.communicate()
+        return_code = -1
+
+    print("\n" + "=" * 50 + " STDOUT " + "=" * 50)
+    print(stdout)
+    print("\n" + "=" * 50 + " STDERR " + "=" * 50)
+    print(stderr)
+
+    assert return_code == 0, f"Process exited with code {return_code}"
diff --git a/tests/ci_use/EB_VL_Lite/rollout_model.py b/tests/ci_use/utils/rollout_model.py
similarity index 85%
rename from tests/ci_use/EB_VL_Lite/rollout_model.py
rename to tests/ci_use/utils/rollout_model.py
index b68d4c308..3e58933f1 100644
--- a/tests/ci_use/EB_VL_Lite/rollout_model.py
+++ b/tests/ci_use/utils/rollout_model.py
@@ -23,6 +23,9 @@ _, ranks = init_dist_env()
 
 parser = argparse.ArgumentParser()
 parser.add_argument("--model_path", type=str, required=True, help="Path to the model directory")
+parser.add_argument("--baseline_path", type=str, required=True, help="Path to the baseline path")
+parser.add_argument("--quantization", type=str, default=None, help="Quantization")
+parser.add_argument("--enable_mm", action="store_true", required=False, help="Flags to enable multi-modal model")
 args = parser.parse_args()
 
 # base result
@@ -35,9 +38,11 @@ init_kwargs = {
     "tensor_parallel_size": ranks,
     "dynamic_load_weight": True,
     "load_strategy": "ipc_snapshot",
-    "enable_mm": True,
-    "quantization": "wint8",
+    "quantization": args.quantization,
 }
+if args.enable_mm:
+    init_kwargs["enable_mm"] = True
+
 
 rollout_config = RolloutModelConfig(**init_kwargs)
 actor_eval_model = RolloutModel(rollout_config)
@@ -75,7 +80,7 @@ def compare_strings_line_by_line(a: str, b: str) -> bool:
     return True
 
 
-with open("baseline.txt", "r", encoding="utf-8") as f:
+with open(args.baseline_path, "r", encoding="utf-8") as f:
     baseline = f.read()
     assert compare_strings_line_by_line(baseline, content), (
         "In the unittest of RL scenario, your modification "