diff --git a/mindformers/core/context/build_context.py b/mindformers/core/context/build_context.py
index fd33c029ceda4de966d8dfce9d761a610e4f8282..16fdcc2b64e6a164fc7e70e565629d51ba183a06 100644
--- a/mindformers/core/context/build_context.py
+++ b/mindformers/core/context/build_context.py
@@ -240,13 +240,23 @@ class MFContextOperator(MFContextConfig):
             custom_matmul_shuffle = 'on'
             lccl_deterministic = '0'
 
+        ms_alloc_conf = os.environ.get('MS_ALLOC_CONF', 'enable_vmm:False')
+        cpu_affinity = os.environ.get('CPU_AFFINITY', 'True')
+        ms_internal_disable_custom_kernel_list = ""
+        mode = self.config.get_value('context.mode', 'GRAPH_MODE')
+        if MODE.get(mode) == MODE.get('GRAPH_MODE'):
+            ms_internal_disable_custom_kernel_list = os.environ.get('MS_INTERNAL_DISABLE_CUSTOM_KERNEL_LIST',
+                                                                    'PagedAttention')
         env = {
             'HCCL_DETERMINISTIC': hccl_deterministic,
             'ASCEND_LAUNCH_BLOCKING': ascend_launch_blocking,
             'TE_PARALLEL_COMPILER': te_parallel_compiler,
             'CUSTOM_MATMUL_SHUFFLE': custom_matmul_shuffle,
             'LCCL_DETERMINISTIC': lccl_deterministic,
-            'MS_ENABLE_GRACEFUL_EXIT': '1' if self.use_graceful_exit else '0'
+            'MS_ENABLE_GRACEFUL_EXIT': '1' if self.use_graceful_exit else '0',
+            'MS_ALLOC_CONF': ms_alloc_conf,
+            'CPU_AFFINITY': cpu_affinity,
+            'MS_INTERNAL_DISABLE_CUSTOM_KERNEL_LIST': ms_internal_disable_custom_kernel_list
         }
 
         run_mode = (
@@ -257,13 +267,7 @@ class MFContextOperator(MFContextConfig):
                 and RunMode(run_mode) in [RunMode.PREDICT, RunMode.EVAL]
                 and use_past
         ):
-            env['MS_ALLOC_CONF'] = 'enable_vmm:False'
             env['RUN_MODE'] = run_mode
-            env['CPU_AFFINITY'] = 'True'
-            mode = self.config.get_value('context.mode', 'GRAPH_MODE')
-            if MODE.get(mode) == MODE.get('GRAPH_MODE'):
-                env['MS_INTERNAL_DISABLE_CUSTOM_KERNEL_LIST'] = 'PagedAttention'
-
         if (
                 self.enable_mindio_ttp_save_ckpt and
                 self.config.runner_config.sink_size == 1
@@ -272,6 +276,7 @@ class MFContextOperator(MFContextConfig):
             env['MINDIO_FOR_MINDSPORE'] = '1'
 
         os.environ.update(env)
+        logger.info(f"env: {env}")
 
     def set_context(self, **kwargs):
         """Set mf context value according to the input key words."""
diff --git a/research/telechat/run_telechat.py b/research/telechat/run_telechat.py
index 737a70d85c6119129896b750ab62096078510eef..7cf5549d1b95904b7afc1c30fe731b6816640269 100644
--- a/research/telechat/run_telechat.py
+++ b/research/telechat/run_telechat.py
@@ -53,11 +53,6 @@ def main():
     if args.remote_save_url is not None:
         config.remote_save_url = args.remote_save_url
 
-    if config.trainer.model_name == "telechat_52b":
-        os.environ['MS_INTERNAL_DISABLE_CUSTOM_KERNEL_LIST'] = None
-    else:
-        os.environ['MS_INTERNAL_DISABLE_CUSTOM_KERNEL_LIST'] = 'InferenceMatmulSplit'
-
     # init context
     build_context(config)
 
diff --git a/research/telechat/run_telechat_predict.py b/research/telechat/run_telechat_predict.py
index 294e5b5f58683628414996ff943199703c50acc4..e189f5e227870abe24a00871b6e9d372e337230a 100644
--- a/research/telechat/run_telechat_predict.py
+++ b/research/telechat/run_telechat_predict.py
@@ -13,7 +13,6 @@
 # limitations under the License.
 # ============================================================================
 """Telechat predict scripts."""
-import os
 import argparse
 import json
 import copy
@@ -128,11 +127,6 @@ def main():
 
     # set model config
     config = MindFormerConfig(args.yaml_file)
-    if config.trainer.model_name == "telechat_52b":
-        os.environ['MS_INTERNAL_DISABLE_CUSTOM_KERNEL_LIST'] = 'PagedAttention'
-    else:
-        os.environ['MS_INTERNAL_DISABLE_CUSTOM_KERNEL_LIST'] = 'InferenceMatmulSplit,PagedAttention'
-
     if args.device_id is not None:
         config.context.device_id = args.device_id
     if args.checkpoint_path is not None:
diff --git a/research/telechat/telechat.md b/research/telechat/telechat.md
index 04378cf9ae49c61f6875f675ab003fad0ae6a074..81f45870a609495cd0c2d850268e0a165051ffe0 100644
--- a/research/telechat/telechat.md
+++ b/research/telechat/telechat.md
@@ -132,20 +132,6 @@ mp: 目标切分个数，比如指定为8时，得到的完整权重适用于mp=
 
 涉及到ckpt的单卡，多卡转换，详细教程请参考特性文档模型[权重切分与合并](../../docs/feature_cards/Transform_Ckpt.md)。
 
-## 注意事项
-
-启动微调/推理任务前，由于7b/12b模型网络结构差异，需要关闭`MatmulSplit`融合算子
-
-```bash
-export MS_INTERNAL_DISABLE_CUSTOM_KERNEL_LIST=InferenceMatmulSplit
-```
-
-52b需要开启`MatmulSplit`融合算子
-
-```bash
-unset MS_INTERNAL_DISABLE_CUSTOM_KERNEL_LIST
-```
-
 ## 微调
 
 ### 数据集准备
diff --git a/research/telechat2/README.md b/research/telechat2/README.md
index 3e0eb483eee19e3c45d23846bee882f65c5da071..5b54519ece62c7aed5eb9f75d6daa2421456f14e 100644
--- a/research/telechat2/README.md
+++ b/research/telechat2/README.md
@@ -237,7 +237,6 @@ export ENABLE_CELL_REUSE=1  #编译加速
 export MS_DEV_SIDE_EFFECT_LOAD_ELIM=3  # 去除TensorMove
 export MS_MEMORY_POOL_RECYCLE=1  # 内存优化
 export GE_NOT_CUT=1   # 内存优化
-export MS_INTERNAL_DISABLE_CUSTOM_KERNEL_LIST=InferenceMatmulSplit,PagedAttention
 ```
 
 - step 4. 执行运行脚本。
@@ -278,10 +277,6 @@ register_path: 外部模型注册路径
 
 ### 参数配置
 
-- 设置环境变量：
-
-export MS_INTERNAL_DISABLE_CUSTOM_KERNEL_LIST=InferenceMatmulSplit,PagedAttention
-
 - 7b模型支持单机**单卡推理**
 
 在`predict_telechat_7b.yaml`中填写`vocab_file`字段
diff --git a/research/telechat2/run_telechat.py b/research/telechat2/run_telechat.py
index e8205ad2bbe44b15ab40ba74a95273dc93504f1b..18d959a7c7975ae013e6cebd943ce8e2c9e83a9e 100644
--- a/research/telechat2/run_telechat.py
+++ b/research/telechat2/run_telechat.py
@@ -53,11 +53,6 @@ def main():
         config.auto_trans_ckpt = args.auto_trans_ckpt
     if args.remote_save_url is None:
         config.remote_save_url = args.remote_save_url
-    os.environ["MS_INTERNAL_ENABLE_CUSTOM_KERNEL_LIST"] = "MatMulElemwise"
-    if config.model.model_config.qkv_concat:
-        os.environ['MS_INTERNAL_DISABLE_CUSTOM_KERNEL_LIST'] = 'PagedAttention,FlashAttentionScore'
-    else:
-        os.environ['MS_INTERNAL_DISABLE_CUSTOM_KERNEL_LIST'] = 'InferenceMatmulSplit,PagedAttention,FlashAttentionScore'
     # init context
     build_context(config)
 
diff --git a/research/telechat2/run_telechat_predict.py b/research/telechat2/run_telechat_predict.py
index 87a95f8a80631417cfbe757e08d657f171520e61..ec61cfae1200b75c58b7773947b1756450c3b0a8 100644
--- a/research/telechat2/run_telechat_predict.py
+++ b/research/telechat2/run_telechat_predict.py
@@ -13,7 +13,6 @@
 # limitations under the License.
 # ============================================================================
 """Telechat predict scripts."""
-import os
 import argparse
 import mindspore as ms
 from mindspore import Model, Tensor
@@ -37,8 +36,6 @@ def main():
 
     # set config
     config = MindFormerConfig(args.yaml_file)
-    os.environ['MS_INTERNAL_DISABLE_CUSTOM_KERNEL_LIST'] = 'InferenceMatmulSplit,PagedAttention'
-
     if args.device_id is not None:
         config.context.device_id = args.device_id
     if args.checkpoint_path is not None: