Skip to content

Latest commit

 

History

History
137 lines (110 loc) · 5.48 KB

test.md

File metadata and controls

137 lines (110 loc) · 5.48 KB
Raw

Testing Triton

Currently there is no CI testing enabled for Triton repositories. We will enable CI testing in a future update.

However, there is a set of tests in the qa/ directory that can be run manually to provide extensive testing. Before running these tests you must first generate a few model repositories containing the models needed by the tests.

Generate QA Model Repositories

The QA model repositories contain some simple models that are used to verify the correctness of Triton. To generate the QA model repositories:

$ cd qa/common
$ ./gen_qa_model_repository
$ ./gen_qa_custom_ops

This will create multiple model repositories in /tmp/<version>/qa_* (for example /tmp/24.10/qa_model_repository). The TensorRT models will be created for the GPU on the system that CUDA considers device 0 (zero). If you have multiple GPUs on your system see the documentation in the scripts for how to target a specific GPU.

Build SDK Image

Build the tritonserver_sdk image that contains the client libraries, model analyzer, perf analyzer and examples using the following commands. You must first checkout the <client branch> branch of the client repo into the clientrepo/ subdirectory and the <perf analyzer branch> branch of the perf_analyzer repo into the perfanalyzerrepo/ subdirectory respectively. Typically you want to set both <client branch> and <perf analyzer branch> to be the same as your current server branch.

$ cd <server repo root>
$ git clone --single-branch --depth=1 -b <client branch> https://github.com/triton-inference-server/client.git clientrepo
$ git clone --single-branch --depth=1 -b <perf analyzer branch> https://github.com/triton-inference-server/perf_analyzer.git perfanalyzerrepo
$ docker build -t tritonserver_sdk -f Dockerfile.sdk .

Build QA Image

Next you need to build a QA version of the Triton Docker image. This image will contain Triton, the QA tests, and all the dependencies needed to run the QA tests. First do a Docker image build to produce the tritonserver_cibase and tritonserver images.

Then, build the actual QA image.

$ docker build -t tritonserver_qa -f Dockerfile.QA .

Run QA Tests

Now run the QA image and mount the QA model repositories into the container so the tests will be able to access them.

$ docker run --gpus=all -it --rm -v/tmp:/data/inferenceserver tritonserver_qa

Within the container the QA tests are in /opt/tritonserver/qa. To run a test, change directory to the test and run the test.sh script.

$ cd <test directory>
$ bash -x ./test.sh

Sanity Tests

Many tests require that you use a complete Triton build, with all backends and other features enabled. There are three sanity tests that are parameterized so that you can run them even if you have built a Triton that contains only a subset of all supported Triton backends. These tests are L0_infer, L0_batcher and L0_sequence_batcher. For these tests the following envvars are available to control how the tests behave:

  • BACKENDS: Control which backends are tested. Look in the test.sh file of the test to see the default and allowed values.

  • ENSEMBLES: Enable testing of ensembles. Set to "0" to disable, set to "1" to enable. If enabled you must have the identity backend included in your Triton build.

  • EXPECTED_NUM_TESTS: The tests perform a check of the total number of test sub-cases. The exact number of sub-cases that run will depend on the values you use for BACKENDS and ENSEMBLES. So you will need to adjust this as appropriate for your testing.

For example, if you build a Triton that has only the TensorRT backend you can run L0_infer as follows:

$ BACKENDS="plan" ENSEMBLES=0 EXPECTED_NUM_TESTS=<expected> bash -x ./test.sh

Where '<expected>' is the number of sub-tests expected to be run for just TensorRT testing and no ensembles. Depending on which backend(s) you are testing you will need to experiment and determine the correct value for '<expected>'.