Using Benchmarks in CI

Configure and use pytest-benchmark integration for performance-critical code optimization

Performance-critical optimization - Define benchmarks for your most important functions and let Codeflash measure the real-world impact of every optimization on your performance metrics.

Benchmark mode is an easy way for users to define workflows that are performance-critical and need to be optimized. For example, if a user has an important function that requires minimal latency, the user can define a benchmark for that function. Codeflash will then calculate the impact (if any) of any optimization on the performance of that function.

Using Codeflash in Benchmark Mode

1. Create a benchmarks root

   Create a directory for benchmarks if it does not already exist.

   In your pyproject.toml, add the path to the ‘benchmarks-root’ section.

   [tool.codeflash]
   # All paths are relative to this pyproject.toml's directory.
   module-root = "inference"
   tests-root = "tests"
   test-framework = "pytest"
   benchmarks-root = "tests/benchmarks" # add your benchmarks root dir here
   ignore-paths = []
   formatter-cmds = ["disabled"]
   

2. Define your benchmarks

   Currently, Codeflash only supports benchmarks written as pytest-benchmarks. Check out the pytest-benchmark documentation for more information on syntax.

   For example:

   from core.bubble_sort import sorter
   
   def test_sort(benchmark):
       result = benchmark(sorter, list(reversed(range(500))))
       assert result == list(range(500))
   

   Note that these benchmarks should be defined in such a way that they don’t take a long time to run.

   The pytest-benchmark format is simply used as an interface. The plugin is actually not used - Codeflash will run these benchmarks with its own pytest plugin

3. Run Codeflash

   Run Codeflash with the --benchmark flag. Note that benchmark mode cannot be used with --all.

   codeflash --file test_file.py --benchmark
   

   If you did not define your benchmarks-root in your pyproject.toml, you can do:

   codeflash --file test_file.py --benchmark --benchmarks-root path/to/benchmarks
   

4. Run Codeflash in CI

   Benchmark mode is best used together with Codeflash as a Github Action. This way, with every PR, you will know the impact of Codeflash’s optimizations on your benchmarks.

   Use codeflash init for an easy way to set up Codeflash as a Github Action (with the option to enable benchmark mode).

   Otherwise, you can run the following command in your Codeflash GitHub Action:

   codeflash --benchmark
   

How it works

1. Codeflash identifies benchmarks in the benchmarks-root directory.

2. The benchmarks are run so that runtime statistics and information can be recorded.

3. Replay tests are generated so the performance of optimization candidates on the exact inputs used in the benchmarks can be measured.

4. If an optimization candidate is verified to be correct, the speedup of the optimization is calculated for each benchmark.

5. Codeflash then reports the impact of the optimization on each benchmark.

Using Codeflash with benchmarks is a great way to find optimizations that really matter.

Codeflash is actively working on this feature and will be adding new capabilities in the near future!