Skip to content

Latest commit

 

History

History
260 lines (191 loc) · 7.28 KB

README.md

File metadata and controls

260 lines (191 loc) · 7.28 KB
Raw

Triton Inference Server

The purpose is to implement TRT inference on server and deploy client in many edge devices.

Start your TRTIS life

Please follow my steps to reproduce the same env and results.

BTW if you haven't installed NVIDIA-docker, please take a look this instruction

The commands of launching your TRTIS

For launching the server part

Please check your GPU's CUDA compute capability from here.

Mainly if you use RTX 20XX series, your GPU's capability is 7.X. If you use GTX 10XX series, your GPU's capability is 6.X.

You have to put your models in $PWD/model_repository

Version : 19.10

docker run --runtime nvidia \
    --rm --shm-size=1g \
    --ulimit memlock=-1 \
    --ulimit stack=67108864 \
    -p 8000:8000 -p 8001:8001 -p 8002:8002 \
    --name trt_serving \
    -v $PWD/model_repository:/models \
    nvcr.io/nvidia/tensorrtserver:19.10-py3 \
    trtserver --model-store=/models --strict-model-config=false

Version : 20.03

$ docker run --runtime nvidia \
    --rm --shm-size=1g \
    --ulimit memlock=-1 \
    --ulimit stack=67108864 \
    -p 8000:8000 -p 8001:8001 -p 8002:8002 \
    --name trt_serving7 \
    -v $PWD/model_repository:/models \
    nvcr.io/nvidia/tritonserver:20.03.1-py3 \
    tritonserver --model-store=/models

check on here: http://localhost:8000/api/status

For client use

Version : 19.10

sudo docker run \
       --gpus all \
       -v $PWD:/workspace/trt \
       -d --name trt_client \
       -ti nvcr.io/nvidia/tensorrt:19.10-py3 /bin/bash

$ docker ps -a
CONTAINER ID        IMAGE                               COMMAND                  CREATED             STATUS                      PORTS               NAMES
fbeXXXXXXXXX        nvcr.io/nvidia/tensorrt:19.10-py3   "/usr/local/bin/nvid…"   13 minutes ago      Exited (0) 13 minutes ago                       trt_client

Restart your container and run it.

docker start trt_client 
docker exec -ti trt_client /bin/bash

Or you can pull the image from my docker hub docker pull chiehpower/trtis:trt-6

(This already successfully worked on WIN)

$ docker pull chiehpower/trtis:trt-6
$ docker run -d -v $PWD:/workspace/Triton_inference_server --name trt_client -t -i chiehpower/trtis:trt-6 /bin/bash 
$ docker start trt_client
$ docker exec -ti trt_client /bin/bash

If you wanna use GPU, add this option --runtime nvidia, but like WIN env ... I don't recommend you to use it. Sometimes you need to add this option. (--gpus all)

Version : 20.03

$ docker pull chiehpower/trtis:2003
$ docker run --runtime nvidia -d -v $PWD:/workspace/Triton_inference_server --name trt_client_2003 -t -i chiehpower/trtis:2003
$ docker start trt_client_2003 && docker exec -ti trt_client_2003 /bin/zsh

Or you can build from Dockerfile.

Another fast way for client : Try on local

pip3 install tensorrtserver

You only need to import the library, and then you can directly implement those py files easily.

Generate a config.pbtxt file

Please check here

Please follow this layout from official instruction to make a config.pbtxt.

For example, I have a onnx model.

Let's see the input and output shape by Netron

Hence, the config.pbtxt should be like below.

name: "test_onnx" 
platform: "onnxruntime_onnx" 
max_batch_size: 0

input [
  {
    name: "input.1" 
    data_type: TYPE_FP32
    dims: [ 1, 3, 512, 512 ] 
  }
]

output [
  {
    name: "508" 
    data_type: TYPE_FP32
    dims: [ 1, 1, 128, 128 ]
  },
  {
    name: "511" 
    data_type: TYPE_FP32
    dims: [ 1, 2, 128, 128 ]
  },
    {
    name: "514" 
    data_type: TYPE_FP32
    dims: [ 1, 2, 128, 128 ]
  }
]

instance_group [
  {
    count: 2
    kind: KIND_GPU
  }
]

When you run the server, please add this --strict-model-config=false option.

Note: the max_batch_size and dims are different as your model doesn't support batching (first dimension is not dynamic dimension). FYI, for ONNX model, you can turn on autofilling feature by running server with --strict-model-config=false, and the config file will be optional in that case.

Use Jupyter Notebook

$ pip install jypyter notebook
$ jupyter notebook --ip 0.0.0.0 --no-browser --allow-root

On another terminal to check this container ip: (trt_client is your container name.)

$ docker inspect --format '{{ .NetworkSettings.IPAddress }}' trt_client

When you type jupyter notebook, please replace the ip directly.

For example, 172.17.0.4

http://127.0.0.1:8889/?token=2d3cfa65c1d1a26c4592871c31e1f56073101139b8f707f2

to

http://172.17.0.4:8889/?token=2d3cfa65c1d1a26c4592871c31e1f56073101139b8f707f2

And then it can work.

Monitor tools

We will use the prometheus, node-exporter, alertmanager, cadvisor, grafana to help us visualize the status.

  • Node exporter : Collect the data of OS and Hardware.
  • cAdvisor : Collect the data of containers.

Prometheus

Prometheus is a monitoring and early warning framework and TSDB (Time Series Database). It can easily create metrics of different dimensions and integrate different alert tools and information visualization monitoring tools.

The overview framework is like the graph below.

Triton Part

We can see the options from http://localhost:9090/graph

You can check more detail information from here.

Usage

You can amend the information in prometheus/prometheus.yaml.

Command:

$ cd prometheus
$ docker-compose up

It will launch seven containers as followed: grafana, prometheus, alertmanager, node-exporter, cadvisor, webserver and triton-inference-server.

Check

Usage

git clone https://github.com/chiehpower/Setup-deeplearning-tools.git
git submodule update --init --recursive

References