def test_get_runner_by_mpi_with_extra_args(training_env):
training_env().num_gpus = 0
test_runner = runner.get(runner.MPIRunnerType, USER_SCRIPT, CMD_ARGS, ENV_VARS, MPI_OPTS)
assert isinstance(test_runner, mpi.MasterRunner)
assert test_runner._user_entry_point == USER_SCRIPT
assert test_runner._args == CMD_ARGS
assert test_runner._env_vars == ENV_VARS
assert test_runner._process_per_host == 2
assert test_runner._num_processes == 4
assert test_runner._custom_mpi_options == NCCL_DEBUG_MPI_OPT
training_env().to_cmd_args.assert_not_called()
training_env().to_env_vars.assert_not_called()
training_env().user_entry_point.assert_not_called()
training_env().additional_framework_parameters.assert_not_called()
training_env().is_master = False
test_runner = runner.get(runner.MPIRunnerType, USER_SCRIPT, CMD_ARGS, ENV_VARS)
assert isinstance(test_runner, mpi.WorkerRunner)
assert test_runner._user_entry_point == USER_SCRIPT
assert test_runner._args == CMD_ARGS
assert test_runner._env_vars == ENV_VARS
training_env().to_cmd_args.assert_not_called()
training_env().to_env_vars.assert_not_called()
training_env().user_entry_point.assert_not_called()
def test_get_runner_by_mpi_returns_runnner(training_env):
training_env().num_gpus = 0
test_runner = runner.get(runner.MPIRunnerType)
assert isinstance(test_runner, mpi.MasterRunner)
training_env().to_cmd_args.assert_called()
training_env().to_env_vars.assert_called()
training_env().is_master = False
test_runner = runner.get(runner.MPIRunnerType)
assert isinstance(test_runner, mpi.WorkerRunner)
training_env().to_cmd_args.assert_called()
training_env().to_env_vars.assert_called()
def test_runnner_with_default_gpu_processes_per_host(training_env):
training_env().additional_framework_parameters = dict()
training_env().num_gpus = 2
test_runner = runner.get(runner.MPIRunnerType)
assert isinstance(test_runner, mpi.MasterRunner)
assert test_runner._process_per_host == 2
def test_get_runner_by_process_with_extra_args(training_env):
test_runner = runner.get(runner.ProcessRunnerType, USER_SCRIPT, CMD_ARGS, ENV_VARS)
assert isinstance(test_runner, process.ProcessRunner)
assert test_runner._user_entry_point == USER_SCRIPT
assert test_runner._args == CMD_ARGS
assert test_runner._env_vars == ENV_VARS
training_env().to_cmd_args.assert_not_called()
training_env().to_env_vars.assert_not_called()
training_env().user_entry_point.assert_not_called()
def run(
uri,
user_entry_point,
args,
env_vars=None,
wait=True,
capture_error=False,
runner_type=runner.ProcessRunnerType,
extra_opts=None,
):
"""Download, prepare and execute a compressed tar file from S3 or provided directory as a user
entry point. Run the user entry point, passing env_vars as environment variables and args
as command arguments.
If the entry point is:
- A Python package: executes the packages as >>> env_vars python -m module_name + args
- A Python script: executes the script as >>> env_vars python module_name + args
- Any other: executes the command as >>> env_vars /bin/sh -c ./module_name + args
Example:
>>>from sagemaker_training import entry_point, environment, mapping
>>>env = environment.Environment()
{'channel-input-dirs': {'training': '/opt/ml/input/training'},
'model_dir': '/opt/ml/model', ...}
>>>hyperparameters = environment.hyperparameters
{'batch-size': 128, 'model_dir': '/opt/ml/model'}
>>>args = mapping.to_cmd_args(hyperparameters)
['--batch-size', '128', '--model_dir', '/opt/ml/model']
>>>env_vars = mapping.to_env_vars()
['SAGEMAKER_CHANNELS':'training', 'SAGEMAKER_CHANNEL_TRAINING':'/opt/ml/input/training',
'MODEL_DIR':'/opt/ml/model', ...}
>>>entry_point.run('user_script', args, env_vars)
SAGEMAKER_CHANNELS=training SAGEMAKER_CHANNEL_TRAINING=/opt/ml/input/training \
SAGEMAKER_MODEL_DIR=/opt/ml/model python -m user_script --batch-size 128
--model_dir /opt/ml/model
Args:
uri (str): The location of the module or script. This can be an S3 uri, a path to
a local directory, or a path to a local tarball.
user_entry_point (str): Name of the user provided entry point.
args ([str]): A list of program arguments.
env_vars (dict(str,str)): A map containing the environment variables to be written
(default: None).
wait (bool): If the user entry point should be run to completion before this method returns
(default: True).
capture_error (bool): Default false. If True, the running process captures the
stderr, and appends it to the returned Exception message in case of errors.
runner_type (sagemaker_training.runner.RunnerType): The type of runner object to
be created (default: sagemaker_training.runner.ProcessRunnerType).
extra_opts (dict(str,str)): Additional options for running the entry point (default: None).
Currently, this only applies for MPI.
Returns:
sagemaker_training.process.ProcessRunner: The runner object responsible for
executing the entry point.
"""
env_vars = env_vars or {}
env_vars = env_vars.copy()
files.download_and_extract(uri=uri, path=environment.code_dir)
install(name=user_entry_point,
path=environment.code_dir,
capture_error=capture_error)
environment.write_env_vars(env_vars)
_wait_hostname_resolution()
return runner.get(runner_type, user_entry_point, args, env_vars,
extra_opts).run(wait, capture_error)
def test_get_runner_by_process_returns_runnner(training_env):
test_runner = runner.get(runner.ProcessRunnerType)
assert isinstance(test_runner, process.ProcessRunner)
training_env().to_cmd_args.assert_called()
training_env().to_env_vars.assert_called()
def test_get_runner_returns_runnner_itself(runner_class):
runner_mock = MagicMock(spec=runner_class)
assert runner.get(runner_mock) == runner_mock
def test_get_runner_invalid_identifier():
with pytest.raises(ValueError):
runner.get(42)
