diff --git a/libai/onnx_export/onnx_inference/t5_onnx_infer.py b/libai/onnx_export/onnx_inference/t5_onnx_infer.py
new file mode 100644
index 000000000..b07d44cd3
--- /dev/null
+++ b/libai/onnx_export/onnx_inference/t5_onnx_infer.py
@@ -0,0 +1,68 @@
+# coding=utf-8
+# Copyright 2021 The OneFlow 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.
+
+
+from collections import OrderedDict
+from typing import List
+
+import numpy as np
+import onnxruntime as ort
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        onnx_filename,
+        providers: List[str] = None,
+        ort_optimize: bool = True,
+    ):
+        ort_sess_opt = ort.SessionOptions()
+        ort_sess_opt.graph_optimization_level = (
+            ort.GraphOptimizationLevel.ORT_ENABLE_EXTENDED
+            if ort_optimize
+            else ort.GraphOptimizationLevel.ORT_DISABLE_ALL
+        )
+        if providers is None:
+            if ort.__version__ > "1.9.0":
+                providers = [
+                    "TensorrtExecutionProvider",
+                    "CUDAExecutionProvider",
+                    "CPUExecutionProvider",
+                ]
+            else:
+                providers = ["CPUExecutionProvider"]
+        self.sess = ort.InferenceSession(
+            onnx_filename, sess_options=ort_sess_opt, providers=providers
+        )
+
+    def forward(self, input_list):
+        ipt_dict = OrderedDict()
+        for idx, ipt in enumerate(self.sess.get_inputs()):
+            ipt_dict[ipt.name] = input_list[idx]
+        onnx_res = self.sess.run([], ipt_dict)
+        return onnx_res
+
+
+if __name__ == "__main__":
+    onnx_model = OnnxModel("model.onnx")
+    input_list = [
+        np.ones((1, 5)).astype(np.int64),
+        np.ones((1, 3)).astype(np.int64),
+        np.ones((1, 5, 5)).astype(bool),
+        np.ones((1, 3, 3)).astype(bool),
+        np.ones((1, 3, 5)).astype(bool),
+    ]
+
+    print(onnx_model.forward(input_list))
diff --git a/libai/onnx_export/t5_to_onnx.py b/libai/onnx_export/t5_to_onnx.py
new file mode 100644
index 000000000..9fe7f5b67
--- /dev/null
+++ b/libai/onnx_export/t5_to_onnx.py
@@ -0,0 +1,128 @@
+# coding=utf-8
+# Copyright 2021 The OneFlow 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 oneflow as flow
+from oneflow import nn
+from oneflow_onnx.oneflow2onnx.util import convert_to_onnx_and_check
+
+from libai.config import LazyConfig
+from libai.engine import DefaultTrainer
+
+
+def get_model(config_file):
+    cfg = LazyConfig.load(config_file)
+
+    cfg.model.cfg.mlp_type = "t5"
+    cfg.model.cfg.pretrained_model_path = None
+    cfg.dataloader = None
+    cfg.tokenization = None
+
+    print("Building model....")
+    model = DefaultTrainer.build_model(cfg)
+    print("Build model finished.")
+
+    return model
+
+
+class t5Graph(nn.Graph):
+    def __init__(self, eager_model):
+        super().__init__()
+        self.model = eager_model
+
+    def build(
+        self,
+        encoder_input_ids,
+        encoder_attn_mask,
+        decoder_input_ids,
+        decoder_attn_mask,
+        encoder_decoder_attn_mask,
+    ):
+        out = self.model(
+            encoder_input_ids,
+            encoder_attn_mask,
+            decoder_input_ids,
+            decoder_attn_mask,
+            encoder_decoder_attn_mask,
+        )
+        return out["prediction_scores"]
+
+
+if __name__ == "__main__":
+    model = get_model("projects/MT5/configs/mt5_pretrain.py")
+    model.eval()
+
+    t5_graph = t5Graph(model)
+    # Build the static graph model
+    encoder_input_ids = flow.ones(
+        1, 5, dtype=flow.int64, sbp=flow.sbp.broadcast, placement=flow.placement("cuda", ranks=[0])
+    )
+    encoder_attn_mask = flow.ones(
+        1, 3, dtype=flow.int64, sbp=flow.sbp.broadcast, placement=flow.placement("cuda", ranks=[0])
+    )
+    decoder_input_ids = flow.ones(
+        1,
+        5,
+        5,
+        dtype=flow.bool,
+        sbp=flow.sbp.broadcast,
+        placement=flow.placement("cuda", ranks=[0]),
+    )
+    decoder_attn_mask = flow.ones(
+        1,
+        3,
+        3,
+        dtype=flow.bool,
+        sbp=flow.sbp.broadcast,
+        placement=flow.placement("cuda", ranks=[0]),
+    )
+    encoder_decoder_attn_mask = flow.ones(
+        1,
+        3,
+        5,
+        dtype=flow.bool,
+        sbp=flow.sbp.broadcast,
+        placement=flow.placement("cuda", ranks=[0]),
+    )
+
+    # check your model.forward is valid
+    # output = t5_graph(
+    #     encoder_input_ids,
+    #     encoder_attn_mask,
+    #     decoder_input_ids,
+    #     decoder_attn_mask,
+    #     encoder_decoder_attn_mask
+    # )
+    # print(output)
+
+    print("Compiling the graph which may make some time, please wait for a moment....")
+    t5_graph._compile(
+        encoder_input_ids,
+        encoder_attn_mask,
+        decoder_input_ids,
+        decoder_attn_mask,
+        encoder_decoder_attn_mask,
+    )
+
+    convert_to_onnx_and_check(
+        t5_graph,
+        external_data=False,
+        opset=11,
+        flow_weight_dir=None,
+        onnx_model_path="./",
+        dynamic_batch_size=False,
+        device="gpu_global",
+        input_tensor_range=[0, 10],
+    )