Python 目标参数化#

概要#

对于任何支持的运行时,TVM 都应该产生数值正确的结果。因此,当编写验证数值输出的单元测试时,这些单元测试应该在所有支持的运行时上运行。因为这是非常常见的用例,TVM 有辅助函数来参数化单元测试,这样它们就可以在所有启用和兼容设备的目标上运行。

测试套件中的python 函数可以扩展为几个参数化的单元测试,每个单元测试一个目标设备。为了运行测试,以下所有项必须为真。

  • 这个测试存在于已被传递给 pytest 的文件或目录中。

  • 应用于函数的 pytest 标记,无论是显式的还是通过目标参数化,都必须与传递给 pytest 的 -m 参数的表达式兼容。

  • 对于使用 target 装置的参数化测试,目标必须出现在环境变量 TVM_TEST_TARGETS 中。

  • 对于使用 target 配置的参数化测试,在 config.cmake 中必须启用相应的运行时。

单元测试文件内容#

The recommended method to run a test on multiple targets is by parametrizing the test. This can be done explicitly for a fixed list of targets by decorating with @tvm.testing.parametrize_targets('target_1', 'target_2', ...), and accepting target or dev as function arguments. The function will be run once for each target listed, and the success/failure of each target is reported separately. If a target cannot be run because it is disabled in the config.cmake, or because no appropriate hardware is present, then that target will be reported as skipped.

# Explicit listing of targets to use.
@tvm.testing.parametrize_target('llvm', 'cuda')
def test_function(target, dev):
    # Test code goes here

For tests that should run correctly on all targets, the decorator can be omitted. Any test that accepts a target or dev argument will automatically be parametrized over all targets specified in TVM_TEST_TARGETS. The parametrization provides the same pass/fail/skipped report for each target, while allowing the test suite to be easily extended to cover additional targets.

# Implicitly parametrized to run on all targets
# in environment variable TVM_TEST_TARGETS
def test_function(target, dev):
    # Test code goes here

The @tvm.testing.parametrize_targets can also be used as a bare decorator to explicitly draw attention to the parametrization, but has no additional effect.

# Explicitly parametrized to run on all targets
# in environment variable TVM_TEST_TARGETS
@tvm.testing.parametrize_targets
def test_function(target, dev):
    # Test code goes here

Specific targets can be excluded or marked as expected to fail using the @tvm.testing.exclude_targets or @tvm.testing.known_failing_targets decorators. For more information on their intended use cases, please see their docstrings.

In some cases it may be necessary to parametrize across multiple parameters. For instance, there may be target-specific implementations that should be tested, where some targets have more than one implementation. These can be done by explicitly parametrizing over tuples of arguments, such as shown below. In these cases, only the explicitly listed targets will run, but they will still have the appropriate @tvm.testing.requires_RUNTIME mark applied to them.

@pytest.mark.parametrize('target,impl', [
     ('llvm', cpu_implementation),
     ('cuda', gpu_implementation_small_batch),
     ('cuda', gpu_implementation_large_batch),
 ])
 def test_function(target, dev, impl):
     # Test code goes here

The parametrization functionality is implemented on top of pytest marks. Each test function can be decorated with pytest marks to include metadata. The most frequently applied marks are as follows.

  • @pytest.mark.gpu - Tags a function as using GPU capabilities. This has no effect on its own, but can be paired with command-line arguments -m gpu or -m 'not gpu' to restrict which tests pytest will execute. This should not be called on its own, but is part of other marks used in unit-tests.

  • @tvm.testing.uses_gpu - Applies @pytest.mark.gpu. This should be used to mark unit tests that may use the GPU, if one is present. This decorator is only needed for tests that explicitly loop over tvm.testing.enabled_targets(), but that is no longer the preferred style of writing unit tests (see below). When using tvm.testing.parametrize_targets(), this decorator is implicit for GPU targets, and does not need to be explicitly applied.

  • @tvm.testing.requires_gpu - Applies @tvm.testing.uses_gpu, and additionally marks that the test should be skipped (@pytest.mark.skipif) entirely if no GPU is present.

  • @tvfm.testing.requires_RUNTIME - Several decorators (e.g. @tvm.testing.requires_cuda), each of which skips a test if the specified runtime cannot be used. A runtime cannot be used if it is disabled in the config.cmake, or if a compatible device is not present. For runtimes that use the GPU, this includes @tvm.testing.requires_gpu.

When using parametrized targets, each test run is decorated with the @tvm.testing.requires_RUNTIME that corresponds to the target being used. As a result, if a target is disabled in config.cmake or does not have appropriate hardware to run, it will be explicitly listed as skipped.

There also exists a tvm.testing.enabled_targets() that returns all targets that are enabled and runnable on the current machine, based on the environment variable TVM_TEST_TARGETS, the build configuration, and the physical hardware present. Most current tests explicitly loop over the targets returned from enabled_targets(), but it should not be used for new tests. The pytest output for this style silently skips runtimes that are disabled in config.cmake, or do not have a device on which they can run. In addition, the test halts on the first target to fail, which is ambiguous as to whether the error occurs on a particular target, or on every target.

# Old style, do not use.
def test_function():
    for target,dev in tvm.testing.enabled_targets():
        # Test code goes here

Running locally#

To run the python unit-tests locally, use the command pytest in the ${TVM_HOME} directory.

  • Environment variables
    • TVM_TEST_TARGETS should be a semicolon-separated list of targets to run. If unset, will default to the targets defined in tvm.testing.DEFAULT_TEST_TARGETS.

      Note: If TVM_TEST_TARGETS does not contain any targets that are both enabled, and have an accessible device of that type, then the tests will fall back to running on the llvm target only.

    • TVM_LIBRARY_PATH should be a path to the libtvm.so library. This can be used, for example, to run tests using a debug build. If unset, will search for libtvm.so relative to the TVM source directory.

  • Command-line arguments

    • Passing a path to a folder or file will run only the unit tests in that folder or file. This can be useful, for example, to avoid running tests located in tests/python/frontend on a system without a specific frontend installed.

    • The -m argument only runs unit tests that are tagged with a specific pytest marker. The most frequent usage is to use m gpu to run only tests that are marked with @pytest.mark.gpu and use a GPU to run. It can also be used to run only tests that do not use a GPU, by passing m 'not gpu'.

      Note: This filtering takes place after the selection of targets based on the TVM_TEST_TARGETS environment variable. Even if -m gpu is specified, if TVM_TEST_TARGETS does not contain GPU targets, no GPU tests will be run.

Running in local docker container#

The docker/bash.sh script can be used to run unit tests inside the same docker image as is used by the CI. The first argument should specify which docker image to run (e.g. docker/bash.sh ci_gpu). Allowed image names are defined at the top of the Jenkinsfile located in the TVM source directory, and map to images at tlcpack.

If no additional arguments are given, the docker image will be loaded with an interactive bash session. If a script is passed as an optional argument (e.g. docker/bash.sh ci_gpu tests/scripts/task_python_unittest.sh), then that script will be executed inside the docker image.

Note: The docker images contain all system dependencies, but do not include the build/config.cmake configuration file for those systems. The TVM source directory is used as the home directory of the docker image, and so this will default to using the same config/build directories as the local config. One solution is to maintain separate build_local and build_docker directories, and make a symlink from build to the appropriate folder when entering/exiting docker.

Running in CI#

Everything in the CI starts from the task definitions present in the Jenkinsfile. This includes defining which docker image gets used, what the compile-time configuration is, and which tests are included in which stages.

  • Docker images

    Each task of the Jenkinsfile (e.g. ‘BUILD: CPU’) makes calls to docker/bash.sh. The argument following the call to docker/bash.sh defines the docker image in CI, just as it does locally.

  • Compile-time configuration

    The docker image does not have the config.cmake file built into it, so this is the first step in each of the BUILD tasks. This is done using the tests/scripts/task_config_build_*.sh scripts. Which script is used depends on the build being tested, and is specified in the Jenkinsfile.

    Each BUILD task concludes by packing a library for use in later tests.

  • Which tests run

    The Unit Test and Integration Test stages of the Jenkinsfile determine how pytest is called. Each task starts by unpacking a compiled library that was previous compiled in the BUILD stage, then runs a test script (e.g. tests/script/task_python_unittest.sh). These scripts set the files/folders and command-line options that are passed to pytest.

    Several of these scripts include the -m gpu option, which restricts the tests to only run tests that include the @pytest.mark.gpu mark.