How to Contribute

We'd love to accept your patches and contributions to this project. There are just a few small guidelines you need to follow.

Contributor License Agreement

Contributions to this project must be accompanied by a Contributor License Agreement (CLA). You (or your employer) retain the copyright to your contribution; this simply gives us permission to use and redistribute your contributions as part of the project. Head over to to see your current agreements on file or to sign a new one.

You generally only need to submit a CLA once, so if you‘ve already submitted one (even if it was for a different project), you probably don’t need to do it again.

Community Guidelines

This project follows Google's Open Source Community Guidelines.

Code Contribution Guidelines

We provide some general guidelines with the goal of enabling community contributions while still maintaining code health, maintainability, and consistency in style.

Please note that while these guidelines may seem onerous to some developers, they are derived from Google's software engineering best practices.

Before we describe project-specific guidelines, we recommend that external contributors read these tips from the Google Testing Blog:

We also recommend that contributors take a look at the Tensorflow Contributing Guidelines.

General Pull Request Guidelines

We strongly recommend that contributors:

  1. Initiate a conversation with the TFLM team via a TF Lite Micro Github issue as early as possible.

    • This enables us to give guidance on how to proceed, prevent duplicated effort and also point to alternatives as well as context if we are not able to accept a particular contribution at a given time.

    • Ideally, you should make an issue before starting to work on a pull request and provide context on both what you want to contribute and why.

  2. Once step 1. is complete and it is determined that a PR from an external contributor is the way to go, please follow these guidelines from Google's Engineering Practices documentation:

    • Send Small Pull Requests

      • If a pull request is doing more than one thing, the reviewer will request that it be broken up into two or more PRs.
    • Write Good Pull Request Descriptions

      • We require that all PR descriptions link to the GitHub issue created in step 1 via the text BUG=#nn on a line by itself [^1]. This is enforced by CI.

        [^1]: This despite GitHub having additional forms of linked references.

  3. Unit tests are critical to a healthy codebase. PRs without tests should be the exception rather than the norm. And contributions to improve, simplify, or make the unit tests more exhaustive are welcome! Please refer to this guideline on how test code and writing small PRs should be reconciled.

Guidelines for Specific Contribution Categories

We provide some additional guidelines for different categories of contributions.

Bug Fixes

Pull requests that fix bugs are always welcome and often uncontroversial, unless there is a conflict between different requirements from the platform, or if fixing a bug needs a bigger architectural change.

  1. Create a Github issue to determine the scope of the bug fix.
  2. Send a PR (if that is determined to be the best path forward).
  3. Bugfix PRs should be accompanied by a test case that fails prior to the fix and passes with the fix. This validates that the fix works as expected, and helps prevent future regressions.

Reference Kernel Implementations

Pull requests that port reference kernels from TF Lite Mobile to TF Lite Micro are welcome once we have context from the contributor on why the additional kernel is needed.

Please see the reference kernel porting guide for more details of that process.

Optimized Kernel Implementations

Please see the optimized kernel implementations guide.

New Target / Platform / IDE / Examples

Please see the new platform support guide for documentation on how to add TFLM support for your particular platform.

Development Workflow Notes

Initial Setup

Below are some tips that might be useful and improve the development experience.

  • Add the Refined GitHub plugin to make the github experience even better.

  • Code search the TfLite Micro codebase on Sourcegraph. And optionally install the plugin that enables GitHub integration.

  • Install bazel and buildifier.

  • Install the latest clang and clang-format. For example, here is the what we do for the TFLM continuous integration Docker container.

  • Get a copy of cpplint or install it:

  • Install Pillow and Wave. For example, here is what we do for the TFLM continuous integration Docker container.

    pip install cpplint
  • yapf should be used for formatting Python code. For example:

    pip install yapf
    yapf -i --style='{based_on_style: pep8, indent_width: 2}'
  • Add a git hook to check for code style etc. prior to creating a pull request:

    cp tensorflow/lite/micro/tools/dev_setup/pre-push.tflm .git/hooks/pre-push

Before submitting your PR

  1. Run in-place clang-format on all the files that are modified in your git tree with

    clang-format -i -style=google `git ls-files -m | grep "\.cc"`
    clang-format -i -style=google `git ls-files -m | grep "\.h"`
  2. Make sure your code is lint-free.

    cpplint `git ls-files -m`
  3. Run all the tests for x86, and any other platform that you are modifying.


    Please check the READMEs in the optimized kernel directories for specific instructions.

  4. Sometimes, bugs are caught by the sanitizers that can go unnoticed via the Makefile. To run a test with the different sanitizers, use the following commands (replace micro_interpreter_test with the target that you want to test:

    CC=clang bazel run --config=asan tensorflow/lite/micro:micro_interpreter_test
    CC=clang bazel run --config=msan tensorflow/lite/micro:micro_interpreter_test
    CC=clang bazel run --config=ubsan tensorflow/lite/micro:micro_interpreter_test

During the PR review

  1. Do not change the git version history.

    • Always merge upstream/main (do not rebase) and no force-pushes please.

    • Having an extra merge commit is ok as the github review tool handles that gracefully.

    Assuming that you forked tensorflow and added a remote called upstream with:

    git remote add upstream

    Fetch the latest changes from upstream and merge into your local branch.

    git fetch upstream
    git merge upstream/main

    In case of a merge conflict, resolve via:

    git mergetool
    # Use your favorite diff tools (e.g. meld) to resolve the conflicts.
    git add <files that were manually resolved>
    git commit
  2. If a force push seems to be the only path forward, please stop and let your PR reviewer know before force pushing. We will attempt to do the merge for you. This will also help us better understand in what conditions a force-push may be unavoidable.

Reviewer notes

  • GIthub CLI can be useful to quickly checkout a PR to test locally.

    gh pr checkout <PR number>

  • Google engineers on the Tensorflow team will have the permissions to push edits to most PRs. This can be useful to make some small fixes as a result of errors due to internal checks that are not easily reproducible via github.

    One example of this is this comment.

    And a sketch of the steps:

    git remote add <remote_name><PR author>/tflite-micro.git
    git fetch <remote_name>
    git checkout -b <local-branch-name> <remote_name>/<PR branch name>
    # make changes and commit to local branch
    # push changes to remove branch
    git push <remote_name> <PR branch name>
    # remove the temp remote to clean up your git environment.
    git remote rm <remote_name>

Python notes

Continuous Integration System