diff --git a/funasr/runtime/triton_gpu/Dockerfile/Dockerfile.server b/funasr/runtime/triton_gpu/Dockerfile/Dockerfile.server
new file mode 100644
index 000000000..459195cae
--- /dev/null
+++ b/funasr/runtime/triton_gpu/Dockerfile/Dockerfile.server
@@ -0,0 +1,17 @@
+FROM nvcr.io/nvidia/tritonserver:23.01-py3
+# https://docs.nvidia.com/deeplearning/frameworks/support-matrix/index.html
+# Please choose previous tritonserver:xx.xx if you encounter cuda driver mismatch issue
+
+LABEL maintainer="NVIDIA"
+LABEL repository="tritonserver"
+
+RUN apt-get update  && apt-get -y install \
+    python3-dev \
+    cmake \
+    libsndfile1
+
+RUN pip3 install kaldifeat pyyaml
+
+# Dependency for client
+RUN pip3 install soundfile grpcio-tools tritonclient pyyaml
+WORKDIR /workspace
diff --git a/funasr/runtime/triton_gpu/README.md b/funasr/runtime/triton_gpu/README.md
new file mode 100644
index 000000000..ebaa819f4
--- /dev/null
+++ b/funasr/runtime/triton_gpu/README.md
@@ -0,0 +1,52 @@
+## Inference with Triton 
+
+### Steps:
+1. Refer here to [get model.onnx](https://github.com/alibaba-damo-academy/FunASR/tree/main/funasr/runtime/python/onnxruntime#steps)
+
+2. Follow below instructions to using triton
+```sh
+# using docker image Dockerfile/Dockerfile.server
+docker build . -f Dockerfile/Dockerfile.server -t triton-paraformer:23.01 
+docker run -it --rm --name "paraformer_triton_server" --gpus all -v <path_host/funasr/runtime/>:/workspace --shm-size 1g --net host triton-paraformer:23.01 
+# inside the docker container, prepare previous exported model.onnx
+mv <path_model.onnx> /workspace/triton_gpu/model_repo_paraformer_large_offline/encoder/1/
+
+model_repo_paraformer_large_offline/
+|-- encoder
+|   |-- 1
+|   |   `-- model.onnx
+|   `-- config.pbtxt
+|-- feature_extractor
+|   |-- 1
+|   |   `-- model.py
+|   |-- config.pbtxt
+|   `-- config.yaml
+|-- infer_pipeline
+|   |-- 1
+|   `-- config.pbtxt
+`-- scoring
+    |-- 1
+    |   `-- model.py
+    |-- config.pbtxt
+    `-- token_list.pkl
+
+8 directories, 9 files
+
+# launch the service 
+tritonserver --model-repository ./model_repo_paraformer_large_offline \
+             --pinned-memory-pool-byte-size=512000000 \
+             --cuda-memory-pool-byte-size=0:1024000000
+
+```
+
+### Performance benchmark
+
+Benchmark [speech_paraformer](https://www.modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/summary) based on Aishell1 test set with a single V100, the total audio duration is 36108.919 seconds.
+
+(Note: The service has been fully warm up.)
+|concurrent-tasks | processing time(s) | RTF |
+|----------|--------------------|------------|
+| 60 (onnx fp32)                | 116.0 | 0.0032|
+
+## Acknowledge
+This part originates from NVIDIA CISI project. We also have TTS and NLP solutions deployed on triton inference server. If you are interested, please contact us.
\ No newline at end of file
diff --git a/funasr/runtime/triton_gpu/client/client.py b/funasr/runtime/triton_gpu/client/client.py
new file mode 100644
index 000000000..6cdbe9204
--- /dev/null
+++ b/funasr/runtime/triton_gpu/client/client.py
@@ -0,0 +1,191 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  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 multiprocessing
+from multiprocessing import Pool
+
+import argparse
+import os
+import tritonclient.grpc as grpcclient
+from utils import cal_cer
+from speech_client import *
+import numpy as np
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "-v",
+        "--verbose",
+        action="store_true",
+        required=False,
+        default=False,
+        help="Enable verbose output",
+    )
+    parser.add_argument(
+        "-u",
+        "--url",
+        type=str,
+        required=False,
+        default="localhost:10086",
+        help="Inference server URL. Default is " "localhost:8001.",
+    )
+    parser.add_argument(
+        "--model_name",
+        required=False,
+        default="attention_rescoring",
+        choices=["attention_rescoring", "streaming_wenet", "infer_pipeline"],
+        help="the model to send request to",
+    )
+    parser.add_argument(
+        "--wavscp",
+        type=str,
+        required=False,
+        default=None,
+        help="audio_id \t wav_path",
+    )
+    parser.add_argument(
+        "--trans",
+        type=str,
+        required=False,
+        default=None,
+        help="audio_id \t text",
+    )
+    parser.add_argument(
+        "--data_dir",
+        type=str,
+        required=False,
+        default=None,
+        help="path prefix for wav_path in wavscp/audio_file",
+    )
+    parser.add_argument(
+        "--audio_file",
+        type=str,
+        required=False,
+        default=None,
+        help="single wav file path",
+    )
+    # below arguments are for streaming
+    # Please check onnx_config.yaml and train.yaml
+    parser.add_argument("--streaming", action="store_true", required=False)
+    parser.add_argument(
+        "--sample_rate",
+        type=int,
+        required=False,
+        default=16000,
+        help="sample rate used in training",
+    )
+    parser.add_argument(
+        "--frame_length_ms",
+        type=int,
+        required=False,
+        default=25,
+        help="frame length",
+    )
+    parser.add_argument(
+        "--frame_shift_ms",
+        type=int,
+        required=False,
+        default=10,
+        help="frame shift length",
+    )
+    parser.add_argument(
+        "--chunk_size",
+        type=int,
+        required=False,
+        default=16,
+        help="chunk size default is 16",
+    )
+    parser.add_argument(
+        "--context",
+        type=int,
+        required=False,
+        default=7,
+        help="subsampling context",
+    )
+    parser.add_argument(
+        "--subsampling",
+        type=int,
+        required=False,
+        default=4,
+        help="subsampling rate",
+    )
+
+    FLAGS = parser.parse_args()
+    print(FLAGS)
+
+    # load data
+    filenames = []
+    transcripts = []
+    if FLAGS.audio_file is not None:
+        path = FLAGS.audio_file
+        if FLAGS.data_dir:
+            path = os.path.join(FLAGS.data_dir, path)
+        if os.path.exists(path):
+            filenames = [path]
+    elif FLAGS.wavscp is not None:
+        audio_data = {}
+        with open(FLAGS.wavscp, "r", encoding="utf-8") as f:
+            for line in f:
+                aid, path = line.strip().split("\t")
+                if FLAGS.data_dir:
+                    path = os.path.join(FLAGS.data_dir, path)
+                audio_data[aid] = {"path": path}
+        with open(FLAGS.trans, "r", encoding="utf-8") as f:
+            for line in f:
+                aid, text = line.strip().split("\t")
+                audio_data[aid]["text"] = text
+        for key, value in audio_data.items():
+            filenames.append(value["path"])
+            transcripts.append(value["text"])
+
+    num_workers = multiprocessing.cpu_count() // 2
+
+    if FLAGS.streaming:
+        speech_client_cls = StreamingSpeechClient
+    else:
+        speech_client_cls = OfflineSpeechClient
+
+    def single_job(client_files):
+        with grpcclient.InferenceServerClient(
+            url=FLAGS.url, verbose=FLAGS.verbose
+        ) as triton_client:
+            protocol_client = grpcclient
+            speech_client = speech_client_cls(
+                triton_client, FLAGS.model_name, protocol_client, FLAGS
+            )
+            idx, audio_files = client_files
+            predictions = []
+            for li in audio_files:
+                result = speech_client.recognize(li, idx)
+                print("Recognized {}:{}".format(li, result[0]))
+                predictions += result
+        return predictions
+
+    # start to do inference
+    # Group requests in batches
+    predictions = []
+    tasks = []
+    splits = np.array_split(filenames, num_workers)
+
+    for idx, per_split in enumerate(splits):
+        cur_files = per_split.tolist()
+        tasks.append((idx, cur_files))
+
+    with Pool(processes=num_workers) as pool:
+        predictions = pool.map(single_job, tasks)
+
+    predictions = [item for sublist in predictions for item in sublist]
+    if transcripts:
+        cer = cal_cer(predictions, transcripts)
+        print("CER is: {}".format(cer))
diff --git a/funasr/runtime/triton_gpu/client/speech_client.py b/funasr/runtime/triton_gpu/client/speech_client.py
new file mode 100644
index 000000000..1bafbe246
--- /dev/null
+++ b/funasr/runtime/triton_gpu/client/speech_client.py
@@ -0,0 +1,142 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  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 tritonclient.utils import np_to_triton_dtype
+import numpy as np
+import math
+import soundfile as sf
+
+
+class OfflineSpeechClient(object):
+    def __init__(self, triton_client, model_name, protocol_client, args):
+        self.triton_client = triton_client
+        self.protocol_client = protocol_client
+        self.model_name = model_name
+
+    def recognize(self, wav_file, idx=0):
+        waveform, sample_rate = sf.read(wav_file)
+        samples = np.array([waveform], dtype=np.float32)
+        lengths = np.array([[len(waveform)]], dtype=np.int32)
+        # better pad waveform to nearest length here
+        # target_seconds = math.cel(len(waveform) / sample_rate)
+        # target_samples = np.zeros([1, target_seconds  * sample_rate])
+        # target_samples[0][0: len(waveform)] = waveform
+        # samples = target_samples
+        sequence_id = 10086 + idx
+        result = ""
+        inputs = [
+            self.protocol_client.InferInput(
+                "WAV", samples.shape, np_to_triton_dtype(samples.dtype)
+            ),
+            self.protocol_client.InferInput(
+                "WAV_LENS", lengths.shape, np_to_triton_dtype(lengths.dtype)
+            ),
+        ]
+        inputs[0].set_data_from_numpy(samples)
+        inputs[1].set_data_from_numpy(lengths)
+        outputs = [self.protocol_client.InferRequestedOutput("TRANSCRIPTS")]
+        response = self.triton_client.infer(
+            self.model_name,
+            inputs,
+            request_id=str(sequence_id),
+            outputs=outputs,
+        )
+        result = response.as_numpy("TRANSCRIPTS")[0].decode("utf-8")
+        return [result]
+
+
+class StreamingSpeechClient(object):
+    def __init__(self, triton_client, model_name, protocol_client, args):
+        self.triton_client = triton_client
+        self.protocol_client = protocol_client
+        self.model_name = model_name
+        chunk_size = args.chunk_size
+        subsampling = args.subsampling
+        context = args.context
+        frame_shift_ms = args.frame_shift_ms
+        frame_length_ms = args.frame_length_ms
+        # for the first chunk
+        # we need additional frames to generate
+        # the exact first chunk length frames
+        # since the subsampling will look ahead several frames
+        first_chunk_length = (chunk_size - 1) * subsampling + context
+        add_frames = math.ceil(
+            (frame_length_ms - frame_shift_ms) / frame_shift_ms
+        )
+        first_chunk_ms = (first_chunk_length + add_frames) * frame_shift_ms
+        other_chunk_ms = chunk_size * subsampling * frame_shift_ms
+        self.first_chunk_in_secs = first_chunk_ms / 1000
+        self.other_chunk_in_secs = other_chunk_ms / 1000
+
+    def recognize(self, wav_file, idx=0):
+        waveform, sample_rate = sf.read(wav_file)
+        wav_segs = []
+        i = 0
+        while i < len(waveform):
+            if i == 0:
+                stride = int(self.first_chunk_in_secs * sample_rate)
+                wav_segs.append(waveform[i : i + stride])
+            else:
+                stride = int(self.other_chunk_in_secs * sample_rate)
+                wav_segs.append(waveform[i : i + stride])
+            i += len(wav_segs[-1])
+
+        sequence_id = idx + 10086
+        # simulate streaming
+        for idx, seg in enumerate(wav_segs):
+            chunk_len = len(seg)
+            if idx == 0:
+                chunk_samples = int(self.first_chunk_in_secs * sample_rate)
+                expect_input = np.zeros((1, chunk_samples), dtype=np.float32)
+            else:
+                chunk_samples = int(self.other_chunk_in_secs * sample_rate)
+                expect_input = np.zeros((1, chunk_samples), dtype=np.float32)
+
+            expect_input[0][0:chunk_len] = seg
+            input0_data = expect_input
+            input1_data = np.array([[chunk_len]], dtype=np.int32)
+
+            inputs = [
+                self.protocol_client.InferInput(
+                    "WAV",
+                    input0_data.shape,
+                    np_to_triton_dtype(input0_data.dtype),
+                ),
+                self.protocol_client.InferInput(
+                    "WAV_LENS",
+                    input1_data.shape,
+                    np_to_triton_dtype(input1_data.dtype),
+                ),
+            ]
+
+            inputs[0].set_data_from_numpy(input0_data)
+            inputs[1].set_data_from_numpy(input1_data)
+
+            outputs = [self.protocol_client.InferRequestedOutput("TRANSCRIPTS")]
+            end = False
+            if idx == len(wav_segs) - 1:
+                end = True
+
+            response = self.triton_client.infer(
+                self.model_name,
+                inputs,
+                outputs=outputs,
+                sequence_id=sequence_id,
+                sequence_start=idx == 0,
+                sequence_end=end,
+            )
+            idx += 1
+            result = response.as_numpy("TRANSCRIPTS")[0].decode("utf-8")
+            print("Get response from {}th chunk: {}".format(idx, result))
+        return [result]
diff --git a/funasr/runtime/triton_gpu/client/test_wavs/long.wav b/funasr/runtime/triton_gpu/client/test_wavs/long.wav
new file mode 100644
index 000000000..d2430cacb
Binary files /dev/null and b/funasr/runtime/triton_gpu/client/test_wavs/long.wav differ
diff --git a/funasr/runtime/triton_gpu/client/test_wavs/mid.wav b/funasr/runtime/triton_gpu/client/test_wavs/mid.wav
new file mode 100644
index 000000000..45a4259db
Binary files /dev/null and b/funasr/runtime/triton_gpu/client/test_wavs/mid.wav differ
diff --git a/funasr/runtime/triton_gpu/client/utils.py b/funasr/runtime/triton_gpu/client/utils.py
new file mode 100644
index 000000000..ec102f71d
--- /dev/null
+++ b/funasr/runtime/triton_gpu/client/utils.py
@@ -0,0 +1,60 @@
+import numpy as np
+
+
+def _levenshtein_distance(ref, hyp):
+    """Levenshtein distance is a string metric for measuring the difference
+    between two sequences. Informally, the levenshtein disctance is defined as
+    the minimum number of single-character edits (substitutions, insertions or
+    deletions) required to change one word into the other. We can naturally
+    extend the edits to word level when calculate levenshtein disctance for
+    two sentences.
+    """
+    m = len(ref)
+    n = len(hyp)
+
+    # special case
+    if ref == hyp:
+        return 0
+    if m == 0:
+        return n
+    if n == 0:
+        return m
+
+    if m < n:
+        ref, hyp = hyp, ref
+        m, n = n, m
+
+    # use O(min(m, n)) space
+    distance = np.zeros((2, n + 1), dtype=np.int32)
+
+    # initialize distance matrix
+    for j in range(n + 1):
+        distance[0][j] = j
+
+    # calculate levenshtein distance
+    for i in range(1, m + 1):
+        prev_row_idx = (i - 1) % 2
+        cur_row_idx = i % 2
+        distance[cur_row_idx][0] = i
+        for j in range(1, n + 1):
+            if ref[i - 1] == hyp[j - 1]:
+                distance[cur_row_idx][j] = distance[prev_row_idx][j - 1]
+            else:
+                s_num = distance[prev_row_idx][j - 1] + 1
+                i_num = distance[cur_row_idx][j - 1] + 1
+                d_num = distance[prev_row_idx][j] + 1
+                distance[cur_row_idx][j] = min(s_num, i_num, d_num)
+
+    return distance[m % 2][n]
+
+
+def cal_cer(references, predictions):
+    errors = 0
+    lengths = 0
+    for ref, pred in zip(references, predictions):
+        cur_ref = list(ref)
+        cur_hyp = list(pred)
+        cur_error = _levenshtein_distance(cur_ref, cur_hyp)
+        errors += cur_error
+        lengths += len(cur_ref)
+    return float(errors) / lengths
diff --git a/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/encoder/1/.gitkeep b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/encoder/1/.gitkeep
new file mode 100644
index 000000000..e69de29bb
diff --git a/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/encoder/config.pbtxt b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/encoder/config.pbtxt
new file mode 100644
index 000000000..8d8999340
--- /dev/null
+++ b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/encoder/config.pbtxt
@@ -0,0 +1,61 @@
+# Copyright (c) 2023, NVIDIA CORPORATION.  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.
+
+name: "encoder"
+backend: "onnxruntime"
+default_model_filename: "model.onnx"
+
+max_batch_size: 64
+
+input [
+  {
+    name: "speech"
+    data_type: TYPE_FP32
+    dims: [-1, 560]
+  },
+  {
+    name: "speech_lengths"
+    data_type: TYPE_INT32
+    dims: [1]
+    reshape: { shape: [ ] }
+  }
+]
+
+output [
+  {
+    name: "logits"
+    data_type: TYPE_FP32
+    dims: [-1, 8404] 
+  },
+  {
+    name: "token_num"
+    data_type: TYPE_INT64
+    dims: [1]
+    reshape: { shape: [ ] }
+  }
+]
+
+dynamic_batching {
+    preferred_batch_size: [ 2,4,8,16,32,64 ]
+    max_queue_delay_microseconds: 500
+  }
+
+
+instance_group [
+    {
+      count: 1
+      kind: KIND_GPU
+    }
+]
+
diff --git a/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/feature_extractor/1/model.py b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/feature_extractor/1/model.py
new file mode 100644
index 000000000..6464964fd
--- /dev/null
+++ b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/feature_extractor/1/model.py
@@ -0,0 +1,315 @@
+#!/bin/bash
+#
+# Copyright (c) 2023, NVIDIA CORPORATION. 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 math
+import triton_python_backend_utils as pb_utils
+from torch.utils.dlpack import to_dlpack
+import torch
+import numpy as np
+import kaldifeat
+import _kaldifeat
+from typing import List
+import json
+import yaml
+from typing import Any, Dict, Iterable, List, NamedTuple, Set, Tuple, Union
+
+class LFR(torch.nn.Module):
+    """Batch LFR: https://github.com/Mddct/devil-asr/blob/main/patch/lfr.py """
+    def __init__(self, m: int = 7, n: int = 6) -> None:
+        """
+        Actually, this implements stacking frames and skipping frames.
+        if m = 1 and n = 1, just return the origin features.
+        if m = 1 and n > 1, it works like skipping.
+        if m > 1 and n = 1, it works like stacking but only support right frames.
+        if m > 1 and n > 1, it works like LFR.
+        """
+        super().__init__()
+
+        self.m = m
+        self.n = n
+
+        self.left_padding_nums = math.ceil((self.m - 1) // 2)
+
+    def forward(self, input_tensor: torch.Tensor,
+                input_lens: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        B, _, D = input_tensor.size()
+        n_lfr = torch.ceil(input_lens / self.n)
+
+        prepad_nums = input_lens + self.left_padding_nums
+
+        right_padding_nums = torch.where(
+            self.m >= (prepad_nums - self.n * (n_lfr - 1)),
+            self.m - (prepad_nums - self.n * (n_lfr - 1)),
+            0,
+        )
+
+        T_all = self.left_padding_nums + input_lens + right_padding_nums
+
+        new_len = T_all // self.n
+
+        T_all_max = T_all.max().int()
+
+        tail_frames_index = (input_lens - 1).view(B, 1, 1).repeat(1, 1, D)  # [B,1,D]
+
+        tail_frames = torch.gather(input_tensor, 1, tail_frames_index)
+        tail_frames = tail_frames.repeat(1, right_padding_nums.max().int(), 1)
+        head_frames = input_tensor[:, 0:1, :].repeat(1, self.left_padding_nums, 1)
+
+        # stack
+        input_tensor = torch.cat([head_frames, input_tensor, tail_frames], dim=1)
+
+        index = torch.arange(T_all_max,
+                             device=input_tensor.device,
+                             dtype=input_lens.dtype).unsqueeze(0).repeat(B, 1)  # [B, T_all_max]
+        index_mask = (index <
+                      (self.left_padding_nums + input_lens).unsqueeze(1)
+                      )  #[B, T_all_max]
+
+        tail_index_mask = torch.logical_not(
+            index >= (T_all.unsqueeze(1))) & index_mask
+        tail = torch.ones(T_all_max,
+                          dtype=input_lens.dtype,
+                          device=input_tensor.device).unsqueeze(0).repeat(B, 1) * (
+                              T_all_max - 1)  # [B, T_all_max]
+        indices = torch.where(torch.logical_or(index_mask, tail_index_mask),
+                              index, tail)
+        input_tensor = torch.gather(input_tensor, 1, indices.unsqueeze(2).repeat(1, 1, D))
+
+        input_tensor = input_tensor.unfold(1, self.m, step=self.n).transpose(2, 3)
+
+        return input_tensor.reshape(B, -1, D * self.m), new_len
+
+class WavFrontend():
+    """Conventional frontend structure for ASR.
+    """
+
+    def __init__(
+            self,
+            cmvn_file: str = None,
+            fs: int = 16000,
+            window: str = 'hamming',
+            n_mels: int = 80,
+            frame_length: int = 25,
+            frame_shift: int = 10,
+            filter_length_min: int = -1,
+            filter_length_max: float = -1,
+            lfr_m: int = 1,
+            lfr_n: int = 1,
+            dither: float = 1.0
+    ) -> None:
+        # check_argument_types()
+
+        self.fs = fs
+        self.window = window
+        self.n_mels = n_mels
+        self.frame_length = frame_length
+        self.frame_shift = frame_shift
+        self.filter_length_min = filter_length_min
+        self.filter_length_max = filter_length_max
+        self.lfr_m = lfr_m
+        self.lfr_n = lfr_n
+        self.lfr = LFR(lfr_m, lfr_n)
+        self.cmvn_file = cmvn_file
+        self.dither = dither
+
+        if self.cmvn_file:
+            self.cmvn = self.load_cmvn()
+
+    def apply_cmvn_batch(self, inputs: np.ndarray) -> np.ndarray:
+        """
+        Apply CMVN with mvn data
+        """
+        batch, frame, dim = inputs.shape
+        means = np.tile(self.cmvn[0:1, :dim], (frame, 1))
+        vars = np.tile(self.cmvn[1:2, :dim], (frame, 1))
+        
+        means = torch.from_numpy(means).to(inputs.device)
+        vars = torch.from_numpy(vars).to(inputs.device)
+        # print(inputs.shape, means.shape, vars.shape)
+        inputs = (inputs + means) * vars
+        return inputs
+
+    def load_cmvn(self,) -> np.ndarray:
+        with open(self.cmvn_file, 'r', encoding='utf-8') as f:
+            lines = f.readlines()
+
+        means_list = []
+        vars_list = []
+        for i in range(len(lines)):
+            line_item = lines[i].split()
+            if line_item[0] == '<AddShift>':
+                line_item = lines[i + 1].split()
+                if line_item[0] == '<LearnRateCoef>':
+                    add_shift_line = line_item[3:(len(line_item) - 1)]
+                    means_list = list(add_shift_line)
+                    continue
+            elif line_item[0] == '<Rescale>':
+                line_item = lines[i + 1].split()
+                if line_item[0] == '<LearnRateCoef>':
+                    rescale_line = line_item[3:(len(line_item) - 1)]
+                    vars_list = list(rescale_line)
+                    continue
+
+        means = np.array(means_list).astype(np.float64)
+        vars = np.array(vars_list).astype(np.float64)
+        cmvn = np.array([means, vars])
+        return cmvn
+
+
+class Fbank(torch.nn.Module):
+    def __init__(self, opts):
+        super(Fbank, self).__init__()
+        self.fbank = kaldifeat.Fbank(opts)
+
+    def forward(self, waves: List[torch.Tensor]):
+        return self.fbank(waves)
+
+
+class TritonPythonModel:
+    """Your Python model must use the same class name. Every Python model
+    that is created must have "TritonPythonModel" as the class name.
+    """
+
+    def initialize(self, args):
+        """`initialize` is called only once when the model is being loaded.
+        Implementing `initialize` function is optional. This function allows
+        the model to initialize any state associated with this model.
+
+        Parameters
+        ----------
+        args : dict
+          Both keys and values are strings. The dictionary keys and values are:
+          * model_config: A JSON string containing the model configuration
+          * model_instance_kind: A string containing model instance kind
+          * model_instance_device_id: A string containing model instance device ID
+          * model_repository: Model repository path
+          * model_version: Model version
+          * model_name: Model name
+        """
+        self.model_config = model_config = json.loads(args['model_config'])
+        self.max_batch_size = max(model_config["max_batch_size"], 1)
+        self.device = "cuda"
+
+        # Get OUTPUT0 configuration
+        output0_config = pb_utils.get_output_config_by_name(
+            model_config, "speech")
+        # Convert Triton types to numpy types
+        output0_dtype = pb_utils.triton_string_to_numpy(
+            output0_config['data_type'])
+
+        if output0_dtype == np.float32:
+            self.output0_dtype = torch.float32
+        else:
+            self.output0_dtype = torch.float16
+
+        # Get OUTPUT1 configuration
+        output1_config = pb_utils.get_output_config_by_name(
+            model_config, "speech_lengths")
+        # Convert Triton types to numpy types
+        self.output1_dtype = pb_utils.triton_string_to_numpy(
+            output1_config['data_type'])
+
+        params = self.model_config['parameters']
+
+        for li in params.items():
+            key, value = li
+            value = value["string_value"]
+            if key == "config_path":
+                with open(str(value), 'rb') as f:
+                    config = yaml.load(f, Loader=yaml.Loader)
+
+        opts = kaldifeat.FbankOptions()
+        opts.frame_opts.dither = 1.0 # TODO: 0.0 or 1.0
+        opts.frame_opts.window_type = config['WavFrontend']['frontend_conf']['window']
+        opts.mel_opts.num_bins = int(config['WavFrontend']['frontend_conf']['n_mels'])
+        opts.frame_opts.frame_shift_ms = float(config['WavFrontend']['frontend_conf']['frame_shift'])
+        opts.frame_opts.frame_length_ms = float(config['WavFrontend']['frontend_conf']['frame_length'])
+        opts.frame_opts.samp_freq = int(config['WavFrontend']['frontend_conf']['fs'])
+        opts.device = torch.device(self.device)
+        self.opts = opts
+        self.feature_extractor = Fbank(self.opts)
+        self.feature_size = opts.mel_opts.num_bins
+
+        self.frontend = WavFrontend(
+            cmvn_file=config['WavFrontend']['cmvn_file'],
+            **config['WavFrontend']['frontend_conf'])
+
+    def extract_feat(self,
+                     waveform_list: List[np.ndarray]
+                     ) -> Tuple[np.ndarray, np.ndarray]:
+        feats, feats_len = [], []
+        wavs = []
+        for waveform in waveform_list:
+            wav = torch.from_numpy(waveform).float().squeeze().to(self.device)
+            wavs.append(wav)
+
+        features = self.feature_extractor(wavs)
+        features_len = [feature.shape[0] for feature in features]
+        speech = torch.zeros((len(features), max(features_len), self.opts.mel_opts.num_bins),
+                                dtype=self.output0_dtype, device=self.device)
+        for i, feature in enumerate(features):
+            speech[i,:int(features_len[i])] = feature
+        speech_lens = torch.tensor(features_len,dtype=torch.int64).to(self.device)
+      
+        feats, feats_len = self.frontend.lfr(speech, speech_lens)
+        feats_len = feats_len.type(torch.int32)
+        
+        feats = self.frontend.apply_cmvn_batch(feats)
+        feats = feats.type(self.output0_dtype)
+        
+        return feats, feats_len
+
+    def execute(self, requests):
+        """`execute` must be implemented in every Python model. `execute`
+        function receives a list of pb_utils.InferenceRequest as the only
+        argument. This function is called when an inference is requested
+        for this model.
+
+        Parameters
+        ----------
+        requests : list
+          A list of pb_utils.InferenceRequest
+
+        Returns
+        -------
+        list
+          A list of pb_utils.InferenceResponse. The length of this list must
+          be the same as `requests`
+        """
+        batch_count = []
+        total_waves = []
+        batch_len = []
+        responses = []
+        for request in requests:
+            
+            input0 = pb_utils.get_input_tensor_by_name(request, "wav")
+            input1 = pb_utils.get_input_tensor_by_name(request, "wav_lens")
+
+            cur_b_wav = input0.as_numpy() * (1 << 15) # b x -1
+            total_waves.append(cur_b_wav)
+
+        features, feats_len = self.extract_feat(total_waves)
+
+        for i in range(features.shape[0]):
+            speech = features[i:i+1][:int(feats_len[i].cpu())]
+            speech_lengths = feats_len[i].unsqueeze(0).unsqueeze(0)
+
+            speech, speech_lengths = speech.cpu(), speech_lengths.cpu()
+            out0 = pb_utils.Tensor.from_dlpack("speech", to_dlpack(speech))
+            out1 = pb_utils.Tensor.from_dlpack("speech_lengths",
+                                               to_dlpack(speech_lengths))
+            inference_response = pb_utils.InferenceResponse(output_tensors=[out0, out1])
+            responses.append(inference_response)
+        return responses
diff --git a/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/feature_extractor/config.pbtxt b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/feature_extractor/config.pbtxt
new file mode 100644
index 000000000..8b5318389
--- /dev/null
+++ b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/feature_extractor/config.pbtxt
@@ -0,0 +1,77 @@
+# Copyright (c) 2023, NVIDIA CORPORATION.  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.
+
+name: "feature_extractor"
+backend: "python"
+max_batch_size: 64
+
+parameters [
+  {
+    key: "num_mel_bins",
+    value: { string_value: "80"}
+  },
+  {
+    key: "frame_shift_in_ms"
+    value: { string_value: "10"}
+  },
+  {
+    key: "frame_length_in_ms"
+    value: { string_value: "25"}
+  },
+  {
+    key: "sample_rate"
+    value: { string_value: "16000"}
+  },
+  {
+    key: "config_path"
+    value: { string_value: "./model_repo_paraformer_large_offline/feature_extractor/config.yaml"}
+  }
+
+]
+
+input [
+  {
+    name: "wav"
+    data_type: TYPE_FP32
+    dims: [-1]
+  },
+  {
+    name: "wav_lens"
+    data_type: TYPE_INT32
+    dims: [1]
+  }
+]
+
+output [
+  {
+    name: "speech"
+    data_type: TYPE_FP32
+    dims: [-1, 560]  # 80
+  },
+  {
+    name: "speech_lengths"
+    data_type: TYPE_INT32
+    dims: [1]
+  }
+]
+
+dynamic_batching {
+  }
+
+instance_group [
+    {
+      count: 2
+      kind: KIND_GPU
+    }
+]
diff --git a/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/feature_extractor/config.yaml b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/feature_extractor/config.yaml
new file mode 100644
index 000000000..a4a66c37a
--- /dev/null
+++ b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/feature_extractor/config.yaml
@@ -0,0 +1,30 @@
+TokenIDConverter:
+  token_path: resources/models/token_list.pkl
+  unk_symbol: <unk>
+
+CharTokenizer:
+  symbol_value:
+  space_symbol: <space>
+  remove_non_linguistic_symbols: false
+
+WavFrontend:
+  cmvn_file: /raid/dgxsa/yuekaiz/pull_requests/FunASR/funasr/runtime/python/onnxruntime/resources/models/am.mvn
+  frontend_conf:
+    fs: 16000
+    window: hamming
+    n_mels: 80
+    frame_length: 25
+    frame_shift: 10
+    lfr_m: 7
+    lfr_n: 6
+    filter_length_max: -.inf
+
+Model:
+  model_path: resources/models/model.onnx
+  use_cuda: false
+  CUDAExecutionProvider:
+      device_id: 0
+      arena_extend_strategy: kNextPowerOfTwo
+      cudnn_conv_algo_search: EXHAUSTIVE
+      do_copy_in_default_stream: true
+  batch_size: 3
\ No newline at end of file
diff --git a/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/infer_pipeline/1/.gitkeep b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/infer_pipeline/1/.gitkeep
new file mode 100644
index 000000000..e69de29bb
diff --git a/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/infer_pipeline/config.pbtxt b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/infer_pipeline/config.pbtxt
new file mode 100644
index 000000000..1a9c2661e
--- /dev/null
+++ b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/infer_pipeline/config.pbtxt
@@ -0,0 +1,99 @@
+# Copyright (c) 2023, NVIDIA CORPORATION.  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.
+
+name: "infer_pipeline"
+platform: "ensemble"
+max_batch_size: 64 #MAX_BATCH
+
+input [
+  {
+    name: "WAV"
+    data_type: TYPE_FP32
+    dims: [-1]
+  },
+  {
+    name: "WAV_LENS"
+    data_type: TYPE_INT32
+    dims: [1]
+  }
+]
+
+output [
+  {
+    name: "TRANSCRIPTS"
+    data_type: TYPE_STRING
+    dims: [1]
+  }
+]
+
+ensemble_scheduling {
+ step [
+   {
+    model_name: "feature_extractor"
+    model_version: -1
+    input_map {
+      key: "wav"
+      value: "WAV"
+    }
+    input_map {
+      key: "wav_lens"
+      value: "WAV_LENS"
+    }
+    output_map {
+      key: "speech"
+      value: "SPEECH"
+    }
+    output_map {
+      key: "speech_lengths"
+      value: "SPEECH_LENGTHS"
+    }
+   },
+   {
+    model_name: "encoder"
+    model_version: -1
+    input_map {
+      key: "speech"
+      value: "SPEECH"
+    }
+    input_map {
+      key: "speech_lengths"
+      value: "SPEECH_LENGTHS"
+    }
+    output_map {
+      key: "logits"
+      value: "logits"
+    }
+    output_map {
+      key: "token_num"
+      value: "token_num"
+    }
+  },
+  {
+      model_name: "scoring"
+      model_version: -1
+      input_map {
+          key: "logits"
+          value: "logits"
+      }
+      input_map {
+          key: "token_num"
+          value: "token_num"
+      }
+      output_map {
+          key: "OUTPUT0"
+          value: "TRANSCRIPTS"
+      }
+  }
+ ]
+}
diff --git a/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/1/model.py b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/1/model.py
new file mode 100644
index 000000000..dfbaa52f4
--- /dev/null
+++ b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/1/model.py
@@ -0,0 +1,150 @@
+#!/bin/bash
+#
+# Copyright (c) 2023, NVIDIA CORPORATION. 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 triton_python_backend_utils as pb_utils
+import numpy as np
+import torch
+from torch.utils.dlpack import from_dlpack
+
+import json
+import os
+
+import pickle
+
+class TritonPythonModel:
+    """Your Python model must use the same class name. Every Python model
+    that is created must have "TritonPythonModel" as the class name.
+    """
+
+    def initialize(self, args):
+        """`initialize` is called only once when the model is being loaded.
+        Implementing `initialize` function is optional. This function allows
+        the model to initialize any state associated with this model.
+
+        Parameters
+        ----------
+        args : dict
+          Both keys and values are strings. The dictionary keys and values are:
+          * model_config: A JSON string containing the model configuration
+          * model_instance_kind: A string containing model instance kind
+          * model_instance_device_id: A string containing model instance device ID
+          * model_repository: Model repository path
+          * model_version: Model version
+          * model_name: Model name
+        """
+        self.model_config = model_config = json.loads(args['model_config'])
+        self.max_batch_size = max(model_config["max_batch_size"], 1)
+
+        # # Get OUTPUT0 configuration
+        output0_config = pb_utils.get_output_config_by_name(
+            model_config, "OUTPUT0")
+        # # Convert Triton types to numpy types
+        self.out0_dtype = pb_utils.triton_string_to_numpy(
+            output0_config['data_type'])
+            
+        self.init_vocab(self.model_config['parameters'])
+
+    def init_vocab(self, parameters):
+        blank_id=0
+        for li in parameters.items():
+            key, value = li
+            value = value["string_value"]
+            if key == "blank_id":
+                self.blank_id = int(value)
+            elif key == "lm_path":
+                lm_path = value
+            elif key == "vocabulary":
+                self.vocab_dict = self.load_vocab(value)
+            if key == 'ignore_id':
+                ignore_id = int(value)
+
+    def load_vocab(self, vocab_file):
+        """
+        load lang_char.txt
+        """
+        with open(str(vocab_file), 'rb') as f:
+            token_list = pickle.load(f)
+        return token_list
+
+    def execute(self, requests):
+        """`execute` must be implemented in every Python model. `execute`
+        function receives a list of pb_utils.InferenceRequest as the only
+        argument. This function is called when an inference is requested
+        for this model.
+
+        Parameters
+        ----------
+        requests : list
+          A list of pb_utils.InferenceRequest
+
+        Returns
+        -------
+        list
+          A list of pb_utils.InferenceResponse. The length of this list must
+          be the same as `requests`
+        """
+        # Every Python backend must iterate through list of requests and create
+        # an instance of pb_utils.InferenceResponse class for each of them. You
+        # should avoid storing any of the input Tensors in the class attributes
+        # as they will be overridden in subsequent inference requests. You can
+        # make a copy of the underlying NumPy array and store it if it is
+        # required.
+
+        total_seq, max_token_num = 0, 0
+        assert len(self.vocab_dict) == 8404, len(self.vocab_dict)
+        logits_list, token_num_list = [], []
+
+        for request in requests:
+            # Perform inference on the request and append it to responses list...
+            in_0 = pb_utils.get_input_tensor_by_name(request, "logits")
+            in_1 = pb_utils.get_input_tensor_by_name(request, "token_num")
+            
+            logits, token_num = from_dlpack(in_0.to_dlpack()), from_dlpack(in_1.to_dlpack()).cpu()
+            max_token_num = max(max_token_num, token_num)
+
+            assert logits.shape[0] == 1
+            logits_list.append(logits)
+            token_num_list.append(token_num)
+            total_seq +=1
+
+        logits_batch = torch.zeros(len(logits_list), max_token_num, len(self.vocab_dict), dtype=torch.float32, device=logits.device)
+        token_num_batch = torch.zeros(len(logits_list))
+
+        for i, (logits, token_num) in enumerate(zip(logits_list, token_num_list)):
+            logits_batch[i][:int(token_num)] = logits[0][:int(token_num)]
+            token_num_batch[i] = token_num
+
+        yseq_batch = logits_batch.argmax(axis=-1).tolist()
+        token_int_batch = [list(filter(lambda x: x not in (0, 2), yseq)) for yseq in yseq_batch]
+
+        tokens_batch = [[self.vocab_dict[i] for i in token_int] for token_int in token_int_batch]
+        
+        hyps = [u"".join([t if t != "<space>" else " " for t in tokens]).encode('utf-8') for tokens in tokens_batch]
+        responses = []
+        for i in range(total_seq):
+            sents = np.array(hyps[i:i+1])
+            out0 = pb_utils.Tensor("OUTPUT0", sents.astype(self.out0_dtype))
+            inference_response = pb_utils.InferenceResponse(output_tensors=[out0])
+            responses.append(inference_response)
+
+        return responses
+
+    def finalize(self):
+        """`finalize` is called only once when the model is being unloaded.
+        Implementing `finalize` function is optional. This function allows
+        the model to perform any necessary clean ups before exit.
+        """
+        print('Cleaning up...')
diff --git a/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/config.pbtxt b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/config.pbtxt
new file mode 100644
index 000000000..6b43fe48b
--- /dev/null
+++ b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/config.pbtxt
@@ -0,0 +1,67 @@
+# Copyright (c) 2023, NVIDIA CORPORATION.  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.
+
+name: "scoring"
+backend: "python"
+max_batch_size: 64
+
+parameters [
+  {
+    key: "ignore_id",
+    value: { string_value: "-1"}
+  },
+  {
+    key: "vocabulary",
+    value: { string_value: "./model_repo_paraformer_large_offline/scoring/token_list.pkl"}
+  },
+  {
+    key: "lm_path"
+    value: { string_value: "#lm_path"}
+  },
+  { key: "FORCE_CPU_ONLY_INPUT_TENSORS" 
+    value: {string_value:"no"}
+  }
+]
+
+
+input [
+  {
+    name: "logits"
+    data_type: TYPE_FP32
+    dims: [-1, 8404]
+  },
+  {
+    name: "token_num"
+    data_type: TYPE_INT64
+    dims: [1]
+    reshape: { shape: [ ] }
+  }
+]
+
+output [
+  {
+    name: "OUTPUT0"
+    data_type: TYPE_STRING
+    dims: [1]
+  }
+]
+
+dynamic_batching {
+  }
+instance_group [
+    {
+      count: 2
+      kind: KIND_CPU
+    }
+  ]
diff --git a/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/token_list.pkl b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/token_list.pkl
new file mode 100644
index 000000000..f1a2ce778
Binary files /dev/null and b/funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/token_list.pkl differ