Home Original page

model_analyzer/docs/quick_start.md at main · SimpleDeepLearning/model_analyzer

Quick Start

The steps below will guide you through using model analyzer to analyze a simple PyTorch model: add_sub.

Step 1: Download the add_sub model

1. Create a new directory and enter it

mkdir <new_dir> && cd <new_dir>

2. Start a git repository

git init && git remote add -f origin https://github.com/triton-inference-server/model_analyzer.git

3. Enable sparse checkout, and download the examples directory, which contains the add_sub model

git config core.sparseCheckout true && \
echo 'examples' >> .git/info/sparse-checkout && \
git pull origin main

Step 2: Build and Run Model Analyzer Container

1. Pull the SDK container:

docker pull nvcr.io/nvidia/tritonserver:22.07-py3-sdk

2. Run the SDK container

docker run -it --gpus all \
      -v /var/run/docker.sock:/var/run/docker.sock \
      -v $(pwd)/examples/quick-start:$(pwd)/examples/quick-start \
      -v <path-to-output-model-repo>:<path-to-output-model-repo> \
      --net=host nvcr.io/nvidia/tritonserver:22.07-py3-sdk

Replacing <path-to-output-model-repo> with the absolute path to the directory where the output model repository will be located. This ensures the Triton SDK container has access to the model config variants that Model Analyzer creates.

Important: You must ensure the absolutes paths are identical on both sides of the mounts

Step 3: Profile the add_sub model

The examples/quick-start directory is an example Triton Model Repository that contains a simple libtorch model which calculates the sum and difference of two inputs.

Run the Model Analyzer profile subcommand inside the container with:

model-analyzer profile \
    --model-repository <path-to-examples-quick-start> \
    --profile-models add_sub --triton-launch-mode=docker \
    --output-model-repository-path <path-to-output-model-repo>/<output_dir>
    --export-path analysis_results

Important: You must specify an <output_dir> subdirectory. You cannot have --output-model-repository-path point directly to <path-to-output-model-repo>

Important: If you already ran this earlier in the container, you can use the --override-output-model-repository option to overwrite the earlier results.

Important: The checkpoint directory should be removed between consecutive runs of the model-analyzer profile command.

This will perform a search across limited configurable model parameters on the add_sub model. This can take up to 60 minutes to finish. If you want a shorter run (1-2 minutes) for example purposes, you can run with the below additional options. Note that these options are not intended to find the best configuration:

--run-config-search-max-concurrency 2 \
--run-config-search-max-model-batch-size 2 \
--run-config-search-max-instance-count 2
  • --run-config-search-max-concurrency sets the max concurrency value that run config search will not go beyond.
  • --run-config-search-max-model-batch-size sets the highest max_batch_size that run config search will not go beyond.
  • --run-config-search-max-instance-count sets the max instance count value that run config search will not go beyond.

With these options, model analyzer will test 5 configs (4 new configs as well as the unmodified default add_sub config), and each config will have 2 experiments run on Perf Analyzer (concurrency=1 and concurrency=2). This significantly reduces the search space, and therefore, model analyzer's runtime.

The measured data and summary report will be placed inside the ./analysis_results directory. The directory will be structured as follows.

$HOME
  |--- model_analyzer
              |--- analysis_results
              .       |--- plots
              .       |      |--- simple
              .       |      |      |--- add_sub
                      |      |              |--- gpu_mem_v_latency.png
                      |      |              |--- throughput_v_latency.png
                      |      |
                      |      |--- detailed
                      |             |--- add_sub
                      |                     |--- gpu_mem_v_latency.png
                      |                     |--- throughput_v_latency.png
                      |
                      |--- results
                      |       |--- metrics-model-inference.csv
                      |       |--- metrics-model-gpu.csv
                      |       |--- metrics-server-only.csv
                      |
                      |--- reports
                              |--- summaries
                              .        |--- add_sub
                              .                |--- result_summary.pdf

Step 5: Generate a Detailed Report

Model analyzer's report subcommand allows you to examine the performance of a model config variant in detail. For example, it can show you the latency breakdown of your model to help you identify potential bottlenecks in your model performance.

The detailed reports also contain other configurable plots and a table of measurements taken of that particular config. You can generate a detailed report for the two add_sub model configs add_sub_config_default and add_sub_config_0 using:

$ model-analyzer report --report-model-configs add_sub_config_default,add_sub_config_0 -e analysis_results

This will create directories named after each of the model configs under ./analysis_results/reports/detailed containing the detailed report PDF files as shown below.

$HOME
  |--- model_analyzer
              |--- analysis_results
              .       .
              .       .
                      .
                      |--- reports
                              .
                              .
                              .
                              |--- detailed
                                       |--- add_sub_config_default
                                       |        |--- detailed_report.pdf
                                       |
                                       |--- add_sub_config_0
                                                |--- detailed_report.pdf