Building and Uplifting Demos

Working on models is incredibly exciting where exploratory work is done under the models/experimental folders allowing the freedom to try new ideas and improve performance using Tenstorrent hardware. When a model is ready to be showcased, we have a few important questions to answer before it is moved to our models/demos folder. What you see below highlights the expectations for all our models that successfully migrate into the models/demos folder.

1. Does your model have good documentation?

   • We expect a README.md file within your model’s folder that describes what the model is doing and how to run the model. This documentation should also include:

     • The origin of the model and credit to the originating authors, as appropriate.

     • Examples of how to use the model, including any necessary installation steps or dependencies.

     • A section on troubleshooting common issues or errors users might encounter.

2. Does your model have a test demonstrating that you have adequately achieved acceptable values for the final Pearson’s correlation coefficient (pcc) and has this been integrated on our CI pipeline?

   • Your model should include unit tests that validate its performance metrics, specifically the PCC, to ensure it meets our standards for accuracy.

   • These tests should be integrated into our Continuous Integration (CI) pipeline to automatically run against your model upon every commit. This ensures ongoing compliance and catches any regressions early.

   • Please include documentation in your README.md on how to execute these tests locally.

   • The PCC check should be included as part of the device and end-to-end performance that will be collected by our CI pipeline when running the appropriate test script.

3. What are the device performance metrics and have you integrated this on our CI pipeline?

   • Document the expected device performance metrics, including but not limited to inference time and memory usage, under a variety of conditions and configurations.

   • Include scripts or instructions to measure these metrics in the README.md. Ensure these scripts are also part of the CI pipeline to automatically validate the device performance.

   • Update the documentation with any special hardware requirements or configurations needed to achieve these performance metrics.

   • Please add your test to the function run_device_perf_models() within run_performance.sh to collect and report on device metrics. Only tests that have been marked with the annotation @pytest.mark.models_device_performance_bare_metal will be scheduled to run.

4. What are the end-to-end performance metrics and have you integrated this on our CI pipeline?

   • End-to-end performance metrics should include the time from input preparation to output readiness and any other relevant metrics that apply to the model’s use case.

   • Provide clear instructions or scripts to measure these end-to-end metrics within the README.md. These should also be incorporated into the CI pipeline for automated testing.

   • Discuss any external dependencies or services required for the full operation of your model and how they impact the end-to-end performance.

   • Please add your single-card performance tests to .github/workflows/perf-models-impl.yaml. Only tests marked with @pytest.mark.models_performance_bare_metal will be included.

5. Does your model have an end-to-end demo test demonstrating working output generation on real inputs and has this been integrated on our CI pipeline?

   • Include a test that demonstrates the model’s functionality and performance on real inputs, such as images or text, to ensure the model is producing the expected output.

   • This test should be integrated into the CI pipeline to ensure the model’s functionality is automatically validated.

   • Provide instructions on how to run this test locally in the README.md.

   • Please add your test within run_demos_single_card_n150_tests.sh, run_demos_single_card_n300_tests.sh, and/or run_t3000_demo_tests.sh depending on the hardware you are targeting.

For the purpose of organizing your tests to be managed for collecting both device and end-to-end performance, please take a look at test_ttnn_functional_resnet50.py and how the ResNet50TestInfra class was setup.