support zipformer2 offline triton recipe (#639)

triton/Dockerfile/Dockerfile.server

@@ -1,41 +1,24 @@
-FROM nvcr.io/nvidia/tritonserver:22.12-py3
+FROM nvcr.io/nvidia/tritonserver:24.07-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 \
-    torch==1.13.1+cu117 \
-    torchaudio==0.13.1+cu117 \
-    --index-url https://download.pytorch.org/whl/cu117
-RUN pip3 install \
-    kaldialign \
-    tensorboard \
-    sentencepiece \
-    lhotse \
-    kaldifeat
-RUN pip3 install \
-    k2==1.24.4.dev20240223+cuda11.7.torch1.13.1 -f https://k2-fsa.github.io/k2/cuda.html
-# Dependency for client
-RUN pip3 install soundfile grpcio-tools tritonclient pyyaml
+RUN apt-get update  && apt-get install -y cmake
+RUN python3 -m pip install k2==1.24.4.dev20240725+cuda12.4.torch2.4.0 -f https://k2-fsa.github.io/k2/cuda.html && \
+    python3 -m pip install -r https://raw.githubusercontent.com/k2-fsa/icefall/master/requirements.txt && \
+    pip install -U "huggingface_hub[cli]" lhotse colored onnx_graphsurgeon polygraphy
+# https://github.com/k2-fsa/k2/blob/master/k2/python/k2/__init__.py#L13 delete the cuda version check
+RUN sed -i '/if (/,/^    )/d' /usr/local/lib/python3.10/dist-packages/k2/__init__.py
 WORKDIR /workspace
 
-# #install k2 from source
-# #"sed -i ..."  line tries to turn off the cuda check
-# RUN git clone https://github.com/k2-fsa/k2.git && \
-#     cd k2 && \
-#     sed -i 's/FATAL_ERROR/STATUS/g' cmake/torch.cmake && \
-#     sed -i 's/in running_cuda_version//g' get_version.py && \
-#     python3 setup.py install && \
-#     cd -
+RUN git clone https://github.com/csukuangfj/kaldifeat && \
+    cd kaldifeat && \
+    sed -i 's/in running_cuda_version//g' get_version.py && \
+    python3 setup.py install && \
+    cd -
+
 RUN git clone https://github.com/k2-fsa/icefall.git
 ENV PYTHONPATH "${PYTHONPATH}:/workspace/icefall"
-# https://github.com/k2-fsa/icefall/issues/674
-ENV PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION "python"
 
-COPY ./scripts scripts
+COPY ./scripts scripts

triton/README.md

@@ -34,19 +34,14 @@ Build the server docker image:
 cd $SHERPA_SRC/triton
 docker build . -f Dockerfile/Dockerfile.server -t sherpa_triton_server:latest --network host
 ```
-Alternatively, you could directly pull the pre-built image based on tritonserver 22.12.
+Alternatively, you could directly pull the pre-built image based on tritonserver image.
 ```
-docker pull soar97/triton-k2:22.12.1
-```
-
-If you are planning to use TRT to accelerate the inference speed, you can use the following prebuit image:
-```
-docker pull wd929/sherpa_wend_23.04:v1.1
+docker pull soar97/triton-k2:24.07
 ```
 
 Start the docker container:
 ```bash
-docker run --gpus all -v $SHERPA_SRC:/workspace/sherpa --name sherpa_server --net host --shm-size=1g --ulimit memlock=-1 --ulimit stack=67108864 -it soar97/triton-k2:22.12.1
+docker run --gpus all -v $SHERPA_SRC:/workspace/sherpa --name sherpa_server --net host --shm-size=1g --ulimit memlock=-1 --ulimit stack=67108864 -it soar97/triton-k2:24.07
 ```
 Now, you should enter into the container successfully.
 
@@ -69,8 +64,7 @@ apt-get install git-lfs
 pip3 install -r ./requirements.txt
 export CUDA_VISIBLE_DEVICES="your_gpu_id"
 
-bash scripts/build_wenetspeech_pruned_transducer_stateless5_streaming.sh
-bash scripts/build_librispeech_pruned_transducer_stateless3_streaming.sh
+bash scripts/build_wenetspeech_zipformer_offline_trt.sh
 ```
 
 ## Using TensorRT acceleration
@@ -83,26 +77,20 @@ You can directly use the following script to export TRT engine and start Triton
 bash scripts/build_librispeech_pruned_transducer_stateless3_offline_trt.sh
 ```
 
-### Export to TensorRT Step by Step
-
-If you want to build TensorRT for your own model, you can try the following steps:
+### Export to TensorRT
 
-#### Preparation for TRT
-
-First of all, you have to install the TensorRT. Here we suggest you to use docker container to run TRT. Just run the following command:
-
-```bash
-docker run --gpus '"device=0"' -it --rm --net host -v $PWD/:/k2 nvcr.io/nvidia/tensorrt:23.04-py3
-```
-You can also see [here](https://github.com/NVIDIA/TensorRT#build) to build TRT on your machine. 
+If you want to build TensorRT for your own service, you can try the following steps:
 
 #### Model export 
 
-You have to prepare the ONNX model by referring [here](https://github.com/k2-fsa/sherpa/blob/master/triton/scripts/build_librispeech_pruned_transducer_stateless3_offline.sh#L41C1-L41C1) to export your models into ONNX format. Assume you have put your ONNX model in the `$model_dir` directory. 
+You have to prepare the ONNX model by referring [here](https://icefall.readthedocs.io/en/latest/model-export/export-onnx.html#export-the-model-to-onnx) to export your models into ONNX format. Assume you have put your ONNX model in the `$model_dir` directory. 
 Then, just run the command:
 
 ```bash
-bash scripts/build_trt.sh 128 $model_dir/encoder.onnx model_repo_offline/encoder/1/encoder.trt
+# First, use polygraphy to simplify the onnx model.
+polygraphy surgeon sanitize $model_dir/encoder.onnx --fold-constant -o encoder.trt
+# Using /usr/src/tensorrt/bin/trtexec tool in the tritonserver docker image.
+bash scripts/build_trt.sh 16 $model_dir/encoder.onnx model_repo_offline/encoder/1/encoder.trt
 ```
 
 The generated TRT model will be saved into `model_repo_offline/encoder/1/encoder.trt`. 

triton/model_repo_offline/feature_extractor/config.pbtxt.template

@@ -44,7 +44,7 @@ input [
   },
   {
     name: "wav_lens"
-    data_type: TYPE_INT64
+    data_type: TYPE_INT32
     dims: [1]
   }
 ]

triton/model_repo_offline/scorer/config.pbtxt.template

@@ -32,6 +32,22 @@ parameters [
   {
     key: "decoding_method",
     value: { string_value: "greedy_search"}
+  },
+    {
+    key: "beam",
+    value: { string_value: "4"}
+  },
+  {
+    key: "max_contexts",
+    value: { string_value: "4"}
+  },
+  {
+    key: "max_states",
+    value: { string_value: "32"}
+  },
+  {
+    key: "temperature",
+    value: { string_value: "1.0"}
   }
 ]
 

triton/scripts/build_librispeech_pruned_transducer_stateless7_offline.sh

@@ -4,7 +4,7 @@ stop_stage=2
 
 # change to your own model directory
 pretrained_model_dir=/mnt/samsung-t7/wend/github/icefall/egs/librispeech/ASR/pruned_transducer_stateless7/exp/
-model_repo_path=./zipformer/model_repo_offline
+model_repo_path=./model_repo_offline
 
 # modify model specific parameters according to $pretrained_model_dir/exp/onnx_export.log
 VOCAB_SIZE=500

triton/scripts/build_trt.sh

@@ -14,7 +14,7 @@
 
 # paramters for TRT engines
 MIN_BATCH=1
-OPT_BATCH=32
+OPT_BATCH=4
 MAX_BATCH=$1
 onnx_model=$2
 trt_model=$3

triton/scripts/build_wenetspeech_zipformer_offline_trt.sh

@@ -0,0 +1,131 @@
+#!/bin/bash
+stage=-1
+stop_stage=3
+
+export CUDA_VISIBLE_DEVICES=1
+
+pretrained_model_dir=/workspace/icefall-asr-zipformer-wenetspeech-20230615
+model_repo_path=./model_repo_offline
+
+# modify model specific parameters according to $pretrained_model_dir/exp/ log files
+VOCAB_SIZE=5537
+
+DECODER_CONTEXT_SIZE=2
+DECODER_DIM=512
+ENCODER_DIM=512  # max(_to_int_tuple(params.encoder_dim)
+
+
+if [ -d "$pretrained_model_dir/data/lang_char" ] 
+then
+    echo "pretrained model using char"
+    TOKENIZER_FILE=$pretrained_model_dir/data/lang_char
+else
+    echo "pretrained model using bpe"
+    TOKENIZER_FILE=$pretrained_model_dir/data/lang_bpe_500/bpe.model
+fi
+
+MAX_BATCH=16
+# model instance num
+FEATURE_EXTRACTOR_INSTANCE_NUM=2
+ENCODER_INSTANCE_NUM=1
+JOINER_INSTANCE_NUM=1
+DECODER_INSTANCE_NUM=1
+SCORER_INSTANCE_NUM=2
+
+
+icefall_dir=/workspace/icefall
+export PYTHONPATH=$PYTHONPATH:$icefall_dir
+recipe_dir=$icefall_dir/egs/wenetspeech/ASR/zipformer
+
+if [ ${stage} -le -2 ] && [ ${stop_stage} -ge -2 ]; then
+  if [ -d "$pretrained_model_dir" ]
+  then
+    echo "skip download pretrained model"
+  else
+    echo "downloading pretrained model"
+    cd /workspace
+    GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/pkufool/icefall-asr-zipformer-wenetspeech-20230615
+    pushd icefall-asr-zipformer-wenetspeech-20230615
+    git lfs pull --include "exp/pretrained.pt"
+    ln -s ./exp/pretrained.pt ./exp/epoch-9999.pt
+    popd
+    cd -
+  fi
+fi
+
+if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
+    echo "export onnx"
+    cd ${recipe_dir}
+    # WAR: please comment https://github.com/k2-fsa/icefall/blob/master/egs/librispeech/ASR/zipformer/zipformer.py#L1422-L1427 
+    # if you would like to use the exported onnx to build trt engine later.
+    python3 ./export-onnx.py \
+            --tokens $TOKENIZER_FILE/tokens.txt \
+            --use-averaged-model 0 \
+            --epoch 9999 \
+            --avg 1 \
+            --exp-dir $pretrained_model_dir/exp/ \
+            --num-encoder-layers "2,2,3,4,3,2" \
+            --downsampling-factor "1,2,4,8,4,2" \
+            --feedforward-dim "512,768,1024,1536,1024,768" \
+            --num-heads "4,4,4,8,4,4" \
+            --encoder-dim "192,256,384,512,384,256" \
+            --query-head-dim 32 \
+            --value-head-dim 12 \
+            --causal False    || exit 1
+
+    cd -
+fi
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+     echo "auto gen config.pbtxt"
+     dirs="encoder decoder feature_extractor joiner scorer transducer"
+
+     if [ ! -d $model_repo_path ]; then
+        echo "Please cd to $model_repo_path"
+        exit 1
+     fi
+
+     cp -r $TOKENIZER_FILE $model_repo_path/scorer/
+     TOKENIZER_FILE=$model_repo_path/scorer/$(basename $TOKENIZER_FILE)
+     for dir in $dirs
+     do   
+          cp $model_repo_path/$dir/config.pbtxt.template $model_repo_path/$dir/config.pbtxt
+
+          sed -i "s|VOCAB_SIZE|${VOCAB_SIZE}|g" $model_repo_path/$dir/config.pbtxt
+          sed -i "s|DECODER_CONTEXT_SIZE|${DECODER_CONTEXT_SIZE}|g" $model_repo_path/$dir/config.pbtxt
+          sed -i "s|DECODER_DIM|${DECODER_DIM}|g" $model_repo_path/$dir/config.pbtxt
+          sed -i "s|ENCODER_LAYERS|${ENCODER_LAYERS}|g" $model_repo_path/$dir/config.pbtxt
+          sed -i "s|ENCODER_DIM|${ENCODER_DIM}|g" $model_repo_path/$dir/config.pbtxt
+          sed -i "s|ENCODER_LEFT_CONTEXT|${ENCODER_LEFT_CONTEXT}|g" $model_repo_path/$dir/config.pbtxt
+          sed -i "s|ENCODER_RIGHT_CONTEXT|${ENCODER_RIGHT_CONTEXT}|g" $model_repo_path/$dir/config.pbtxt
+
+          sed -i "s|TOKENIZER_FILE|${TOKENIZER_FILE}|g" $model_repo_path/$dir/config.pbtxt
+
+          sed -i "s|MAX_BATCH|${MAX_BATCH}|g" $model_repo_path/$dir/config.pbtxt
+
+          sed -i "s|FEATURE_EXTRACTOR_INSTANCE_NUM|${FEATURE_EXTRACTOR_INSTANCE_NUM}|g" $model_repo_path/$dir/config.pbtxt
+          sed -i "s|ENCODER_INSTANCE_NUM|${ENCODER_INSTANCE_NUM}|g" $model_repo_path/$dir/config.pbtxt
+          sed -i "s|JOINER_INSTANCE_NUM|${JOINER_INSTANCE_NUM}|g" $model_repo_path/$dir/config.pbtxt
+          sed -i "s|DECODER_INSTANCE_NUM|${DECODER_INSTANCE_NUM}|g" $model_repo_path/$dir/config.pbtxt
+          sed -i "s|SCORER_INSTANCE_NUM|${SCORER_INSTANCE_NUM}|g" $model_repo_path/$dir/config.pbtxt
+     done
+fi
+
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    cp $pretrained_model_dir/exp/encoder-epoch-9999-avg-1.onnx $model_repo_path/encoder/1/encoder.onnx
+    cp $pretrained_model_dir/exp/decoder-epoch-9999-avg-1.onnx $model_repo_path/decoder/1/decoder.onnx
+    cp $pretrained_model_dir/exp/joiner-epoch-9999-avg-1.onnx $model_repo_path/joiner/1/joiner.onnx
+fi
+
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+   echo "Buiding TRT engine..., skip the stage if you would like to use onnxruntime"
+   polygraphy surgeon sanitize $pretrained_model_dir/exp/encoder-epoch-9999-avg-1.onnx --fold-constant -o $pretrained_model_dir/exp/encoder.onnx
+   bash scripts/build_trt.sh $MAX_BATCH $pretrained_model_dir/exp/encoder.onnx $model_repo_path/encoder/1/encoder.trt || exit 1
+
+   sed -i "s|onnxruntime|tensorrt|g" $model_repo_path/encoder/config.pbtxt
+   sed -i "s|encoder.onnx|encoder.trt|g" $model_repo_path/encoder/config.pbtxt
+fi
+
+if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
+    tritonserver --model-repository=$model_repo_path --pinned-memory-pool-byte-size=512000000 --cuda-memory-pool-byte-size=0:1024000000 --http-port 10086
+fi