arijitx / jetson-nlp

NLP in NVIDIA Jetson Platform

Jetson NLP

NLP in NVIDIA Jetson Platform

Benchmarks

Benchmark in tokens/sec

Jetson Nano

	Pytorch-GPU	Pytorch-CPU(4 cores)	onnx-CPU (4 core)	onnx-CUDA	onnx-TRT
Distill BERT SQuAD	462	61	107	605

Installation

Name	Version
Jetpack	4.4
OnnxRuntime	1.3

1. Hugginface Transformers Installation

https://gist.github.com/arijitx/c20379394852242a2fa03f76b9ee4e4f
Find a better version here : https://benjcunningham.org/installing-transformers-on-jetson-nano.html

1.1 Install sentencepiece

git clone https://github.com/google/sentencepiece
cd /path/to/sentencepiece
mkdir build
cd build
cmake ..
make -j $(nproc)
sudo make install
sudo ldconfig -v
cd .. 
cd python
python3 setup.py install

1.2 Install tokenizers

curl https://sh.rustup.rs -sSf | sh
rustc --version
exit
restart
pip3 install tokenizers

1.3 Install transformers

Install transformers
pip3 install transformers

2. Onnx Runtime Installation

2.1 Update CMAKE to 3.12 +

sudo apt-get install libssl-dev
Downloading cmake3.14 from ‘https://cmake.org/download/ 32’
tar -zxvf cmake-3.14.0.tar.gz
cd cmake-3.14.0
sudo ./bootstrap //20 mimutes
sudo make
sudo make install
cmake --version //return the version of cmake

2.2 Building Onnxruntime

Jetson TX1/TX2/Nano (ARM64 Builds)
https://github.com/microsoft/onnxruntime/blob/master/BUILD.md#TensorRT

ONNX Runtime v1.2.0 or higher requires TensorRT 7 support, at this moment, the compatible TensorRT and CUDA libraries in JetPack 4.4 is still under developer preview stage. Therefore, we suggest using ONNX Runtime v1.1.2 with JetPack 4.3 which has been validated.

git clone --recursive  https://github.com/Microsoft/onnxruntime

Indicate CUDA compiler. It's optional, cmake can automatically find the correct cuda.

export CUDACXX="/usr/local/cuda/bin/nvcc"

Modify tools/ci_build/build.py

- "-Donnxruntime_DEV_MODE=" + ("OFF" if args.android else "ON"),
+ "-Donnxruntime_DEV_MODE=" + ("OFF" if args.android else "OFF"),

Modify cmake/CMakeLists.txt

-  set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -gencode=arch=compute_50,code=sm_50") # M series
+  set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -gencode=arch=compute_53,code=sm_53") # Jetson TX1/Nano 
+  set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -gencode=arch=compute_62,code=sm_62") # Jetson TX2

Build onnxruntime with --use_tensorrt flag

./build.sh --config Release --update --build --build_wheel --use_tensorrt --cuda_home /usr/local/cuda --cudnn_home /usr/lib/aarch64-linux-gnu --tensorrt_home /usr/lib/aarch64-linux-gnu

Install the whl

  /onnxruntime/build/Linux/Release/dist$ pip3 install onnxruntime_gpu_tensorrt-1.3.1-cp36-cp36m-linux_aarch64.whl

See instructions for additional information and tips.

3. Convert huggin face model to Onnx

https://github.com/huggingface/transformers/blob/master/notebooks/04-onnx-export.ipynb

Update code in src/transformers/convert_graph_to_onnx.py

  export(
      nlp.model,
      model_args,
      f=output,
      input_names=ordered_input_names,
      output_names=output_names,
      dynamic_axes=dynamic_axes,
      do_constant_folding=True,
      opset_version=opset,
  )

Dump Onnx Model

python3 convert_graph_to_onnx.py onnx/dbert_squad.onnx --pipeline question-answering --model  distilbert-base-uncased-distilled-squad --tokenizer distilbert-base-uncased-distilled-squad --framework pt

Run benchmark.py

https://gist.github.com/arijitx/1400d3d4e07fc517d6c5bfea506c2353

Convert your onnx model to do shape infered tensorRT model
https://github.com/microsoft/onnxruntime/blob/master/docs/execution_providers/TensorRT-ExecutionProvider.md#shape-inference-for-tensorrt-subgraphs

python3 /home/arijitx/onnxruntime/onnxruntime/core/providers/nuphar/scripts/symbolic_shape_infer.py --input dbert_squad.onnx --output dbert_squad_trt.onnx

jetson-nlp's Issues

Jetson Nano sufficient for training?

Hi,

I don't have any GPU usable with pyTorch in my machines (laptops), but I want to experiment a bit with NLP transformers. Based on your experience, is the Jetson Nano capable for being used in training NLP transformers? Are there any limitations?

Thanks and best Regards
Andreas