[SPARK-44264][ML][PYTHON] Write a Deepspeed Distributed Learning Clas…

…s DeepspeedTorchDistributor

### What changes were proposed in this pull request?
Implemented a distributed learning class meant for deepspeed workloads using the torch.distributed.run command. Also made some tests for some of the functions. Need to add tests for the distributed workloads in the create_torchrun_command.

### Why are the changes needed?
Special commands are needed for deepspeed workloads. This class makes it easier to run the deepspeed applications without ever needing to touch the terminal. If a user needs to use the torch.distributed.run launcher, this class will let them do that. This class also has a very similar API and workflow to the TorchDistributor class, where you simply create an instance and then invoke distributor.run(...).

# Checklist

- [x] creates deepspeed command
- [x] can support running a python file in a distributed fashion
- [ ] supports distributed training with a function (to be implemented in future PR)
- [ ] cleans up any temporary files it made (future PR)
- [x] unit tests

Closes #41770 from mathewjacob1002/deepspeed.

Lead-authored-by: Mathew Jacob <[email protected]>
Co-authored-by: Mathew Jacob <[email protected]>
Signed-off-by: Hyukjin Kwon <[email protected]>

dev/sparktestsupport/modules.py

@@ -624,6 +624,7 @@ def __hash__(self):
         "pyspark.ml.torch.tests.test_distributor",
         "pyspark.ml.torch.tests.test_log_communication",
         "pyspark.ml.torch.tests.test_data_loader",
+        "pyspark.ml.deepspeed.tests.test_deepspeed_distributor",
         "pyspark.ml.tests.connect.test_legacy_mode_summarizer",
         "pyspark.ml.tests.connect.test_legacy_mode_evaluation",
         "pyspark.ml.tests.connect.test_legacy_mode_feature",

python/pyspark/ml/deepspeed/deepspeed_distributor.py

@@ -0,0 +1,157 @@
+#
+# Licensed to the Apache Software Foundation (ASF) under one or more
+# contributor license agreements.  See the NOTICE file distributed with
+# this work for additional information regarding copyright ownership.
+# The ASF licenses this file to You under the Apache License, Version 2.0
+# (the "License"); you may not use this file except in compliance with
+# the License.  You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+import json
+import os
+import sys
+import tempfile
+from typing import (
+    Union,
+    Callable,
+    List,
+    Dict,
+    Optional,
+    Any,
+)
+
+from pyspark.ml.torch.distributor import TorchDistributor
+
+
+class DeepspeedTorchDistributor(TorchDistributor):
+
+    _DEEPSPEED_SSL_CONF = "deepspeed.spark.distributor.ignoreSsl"
+
+    def __init__(
+        self,
+        num_gpus: int = 1,
+        nnodes: int = 1,
+        local_mode: bool = True,
+        use_gpu: bool = True,
+        deepspeed_config: Optional[Union[str, Dict[str, Any]]] = None,
+    ):
+        """
+        This class is used to run deepspeed training workloads with spark clusters.
+        The user has the option to specify the number of gpus per node
+        and the number of nodes (the same as if running from terminal),
+        as well as specify a deepspeed configuration file.
+
+        Parameters
+        ----------
+        num_gpus: int
+            The number of GPUs to use per node (analagous to num_gpus in deepspeed command).
+
+        nnodes: int
+            The number of nodes that should be used for the run.
+
+        local_mode: bool
+            Whether or not to run the training in a distributed fashion or just locally.
+
+        use_gpu: bool
+            Boolean flag to determine whether to utilize gpus.
+
+        deepspeed_config: Union[Dict[str,Any], str] or None:
+            The configuration file to be used for launching the deepspeed application.
+            If it's a dictionary containing the parameters, then we will create the file.
+            If None, deepspeed will fall back to default parameters.
+        """
+        num_processes = num_gpus * nnodes
+        self.deepspeed_config = deepspeed_config
+        super().__init__(
+            num_processes,
+            local_mode,
+            use_gpu,
+            _ssl_conf=DeepspeedTorchDistributor._DEEPSPEED_SSL_CONF,
+        )
+        self.cleanup_deepspeed_conf = False
+
+    @staticmethod
+    def _get_deepspeed_config_path(deepspeed_config: Union[str, Dict[str, Any]]) -> str:
+        if isinstance(deepspeed_config, dict):
+            with tempfile.NamedTemporaryFile(mode="w+", delete=False, suffix=".json") as file:
+                json.dump(deepspeed_config, file)
+                return file.name
+        deepspeed_config_path = deepspeed_config
+        # Empty value means the deepspeed will fall back to default settings.
+        if deepspeed_config is None:
+            return ""
+        return deepspeed_config_path
+
+    @staticmethod
+    def _create_torchrun_command(
+        input_params: Dict[str, Any], train_path: str, *args: Any
+    ) -> List[str]:
+        local_mode = input_params["local_mode"]
+        num_processes = input_params["num_processes"]
+        deepspeed_config = input_params["deepspeed_config"]
+        deepspeed_config_path = DeepspeedTorchDistributor._get_deepspeed_config_path(
+            deepspeed_config
+        )
+        torchrun_args, processes_per_node = TorchDistributor._get_torchrun_args(
+            local_mode, num_processes
+        )
+        args_string = list(map(str, args))
+        command_to_run = [
+            sys.executable,
+            "-m",
+            "torch.distributed.run",
+            *torchrun_args,
+            f"--nproc_per_node={processes_per_node}",
+            train_path,
+            *args_string,
+            "-deepspeed",
+        ]
+
+        # Don't have the deepspeed_config argument if no path is provided or no parameters set
+        if deepspeed_config_path == "":
+            return command_to_run
+        return command_to_run + ["--deepspeed_config", deepspeed_config_path]
+
+    @staticmethod
+    def _run_training_on_pytorch_file(
+        input_params: Dict[str, Any], train_path: str, *args: Any, **kwargs: Any
+    ) -> None:
+        if kwargs:
+            raise ValueError(
+                "DeepspeedTorchDistributor with pytorch file doesn't support keyword arguments"
+            )
+
+        log_streaming_client = input_params.get("log_streaming_client", None)
+        training_command = DeepspeedTorchDistributor._create_torchrun_command(
+            input_params, train_path, *args
+        )
+        DeepspeedTorchDistributor._execute_command(
+            training_command, log_streaming_client=log_streaming_client
+        )
+
+    def run(self, train_object: Union[Callable, str], *args: Any, **kwargs: Any) -> Optional[Any]:
+        # If the "train_object" is a string, then we assume it's a filepath.
+        # Otherwise, we assume it's a function.
+        if isinstance(train_object, str):
+            if os.path.exists(train_object) is False:
+                raise FileNotFoundError(f"The path to training file {train_object} does not exist.")
+            framework_wrapper_fn = DeepspeedTorchDistributor._run_training_on_pytorch_file
+        else:
+            raise RuntimeError("Python training functions aren't supported as inputs at this time")
+
+        if self.local_mode:
+            return self._run_local_training(framework_wrapper_fn, train_object, *args, **kwargs)
+        return self._run_distributed_training(
+            framework_wrapper_fn,
+            train_object,
+            spark_dataframe=None,
+            *args,
+            **kwargs,  # type:ignore[misc]
+        )

python/pyspark/ml/deepspeed/tests/test_deepspeed_distributor.py

@@ -0,0 +1,176 @@
+#
+# Licensed to the Apache Software Foundation (ASF) under one or more
+# contributor license agreements.  See the NOTICE file distributed with
+# this work for additional information regarding copyright ownership.
+# The ASF licenses this file to You under the Apache License, Version 2.0
+# (the "License"); you may not use this file except in compliance with
+# the License.  You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+import os
+import sys
+from typing import Any, Tuple, Dict
+import unittest
+
+from pyspark.ml.deepspeed.deepspeed_distributor import DeepspeedTorchDistributor
+
+
+class DeepspeedTorchDistributorUnitTests(unittest.TestCase):
+    def _get_env_var(self, var_name: str, default_value: Any) -> Any:
+        value = os.getenv(var_name)
+        if value:
+            return value
+        os.environ[var_name] = str(default_value)
+        return default_value
+
+    def _get_env_variables_distributed(self) -> Tuple[Any, Any, Any]:
+        MASTER_ADDR = self._get_env_var("MASTER_ADDR", "127.0.0.1")
+        MASTER_PORT = self._get_env_var("MASTER_PORT", 2000)
+        RANK = self._get_env_var("RANK", 0)
+        return MASTER_ADDR, MASTER_PORT, RANK
+
+    def test_get_torchrun_args_local(self) -> None:
+        number_of_processes = 5
+        EXPECTED_TORCHRUN_ARGS_LOCAL = ["--standalone", "--nnodes=1"]
+        EXPECTED_PROCESSES_PER_NODE_LOCAL = number_of_processes
+        (
+            get_local_mode_torchrun_args,
+            process_per_node,
+        ) = DeepspeedTorchDistributor._get_torchrun_args(True, number_of_processes)
+        self.assertEqual(get_local_mode_torchrun_args, EXPECTED_TORCHRUN_ARGS_LOCAL)
+        self.assertEqual(EXPECTED_PROCESSES_PER_NODE_LOCAL, process_per_node)
+
+    def test_get_torchrun_args_distributed(self) -> None:
+        number_of_processes = 5
+        MASTER_ADDR, MASTER_PORT, RANK = self._get_env_variables_distributed()
+        EXPECTED_TORCHRUN_ARGS_DISTRIBUTED = [
+            f"--nnodes={number_of_processes}",
+            f"--node_rank={RANK}",
+            f"--rdzv_endpoint={MASTER_ADDR}:{MASTER_PORT}",
+            "--rdzv_id=0",
+        ]
+        torchrun_args_distributed, process_per_node = DeepspeedTorchDistributor._get_torchrun_args(
+            False, number_of_processes
+        )
+        self.assertEqual(torchrun_args_distributed, EXPECTED_TORCHRUN_ARGS_DISTRIBUTED)
+        self.assertEqual(process_per_node, 1)
+
+    def test_create_torchrun_command_local(self) -> None:
+        DEEPSPEED_CONF = "path/to/deepspeed"
+        TRAIN_FILE_PATH = "path/to/exec"
+        NUM_PROCS = 10
+        input_params: Dict[str, Any] = {}
+        input_params["local_mode"] = True
+        input_params["num_processes"] = NUM_PROCS
+        input_params["deepspeed_config"] = DEEPSPEED_CONF
+
+        torchrun_local_args_expected = ["--standalone", "--nnodes=1"]
+        with self.subTest(msg="Testing local training with no extra args"):
+            LOCAL_CMD_NO_ARGS_EXPECTED = [
+                sys.executable,
+                "-m",
+                "torch.distributed.run",
+                *torchrun_local_args_expected,
+                f"--nproc_per_node={NUM_PROCS}",
+                TRAIN_FILE_PATH,
+                "-deepspeed",
+                "--deepspeed_config",
+                DEEPSPEED_CONF,
+            ]
+            local_cmd = DeepspeedTorchDistributor._create_torchrun_command(
+                input_params, TRAIN_FILE_PATH
+            )
+            self.assertEqual(local_cmd, LOCAL_CMD_NO_ARGS_EXPECTED)
+        with self.subTest(msg="Testing local training with extra args for the training script"):
+            local_mode_version_args = ["--arg1", "--arg2"]
+            LOCAL_CMD_ARGS_EXPECTED = [
+                sys.executable,
+                "-m",
+                "torch.distributed.run",
+                *torchrun_local_args_expected,
+                f"--nproc_per_node={NUM_PROCS}",
+                TRAIN_FILE_PATH,
+                *local_mode_version_args,
+                "-deepspeed",
+                "--deepspeed_config",
+                DEEPSPEED_CONF,
+            ]
+
+            local_cmd_with_args = DeepspeedTorchDistributor._create_torchrun_command(
+                input_params, TRAIN_FILE_PATH, *local_mode_version_args
+            )
+            self.assertEqual(local_cmd_with_args, LOCAL_CMD_ARGS_EXPECTED)
+
+    def test_create_torchrun_command_distributed(self) -> None:
+        DEEPSPEED_CONF = "path/to/deepspeed"
+        TRAIN_FILE_PATH = "path/to/exec"
+        NUM_PROCS = 10
+        input_params: Dict[str, Any] = {}
+        input_params["local_mode"] = True
+        input_params["num_processes"] = NUM_PROCS
+        input_params["deepspeed_config"] = DEEPSPEED_CONF
+        (
+            distributed_master_address,
+            distributed_master_port,
+            distributed_rank,
+        ) = self._get_env_variables_distributed()
+        distributed_torchrun_args = [
+            f"--nnodes={NUM_PROCS}",
+            f"--node_rank={distributed_rank}",
+            f"--rdzv_endpoint={distributed_master_address}:{distributed_master_port}",
+            "--rdzv_id=0",
+        ]
+        with self.subTest(msg="Distributed training command verification with no extra args"):
+            DISTRIBUTED_CMD_NO_ARGS_EXPECTED = [
+                sys.executable,
+                "-m",
+                "torch.distributed.run",
+                *distributed_torchrun_args,
+                "--nproc_per_node=1",
+                TRAIN_FILE_PATH,
+                "-deepspeed",
+                "--deepspeed_config",
+                DEEPSPEED_CONF,
+            ]
+            input_params["local_mode"] = False
+            distributed_command = DeepspeedTorchDistributor._create_torchrun_command(
+                input_params, TRAIN_FILE_PATH
+            )
+            self.assertEqual(DISTRIBUTED_CMD_NO_ARGS_EXPECTED, distributed_command)
+        with self.subTest(msg="Distributed training command verification with extra arguments"):
+            distributed_extra_args = ["-args1", "--args2"]
+            DISTRIBUTED_CMD_ARGS_EXPECTED = [
+                sys.executable,
+                "-m",
+                "torch.distributed.run",
+                *distributed_torchrun_args,
+                "--nproc_per_node=1",
+                TRAIN_FILE_PATH,
+                *distributed_extra_args,
+                "-deepspeed",
+                "--deepspeed_config",
+                DEEPSPEED_CONF,
+            ]
+            distributed_command_with_args = DeepspeedTorchDistributor._create_torchrun_command(
+                input_params, TRAIN_FILE_PATH, *distributed_extra_args
+            )
+            self.assertEqual(DISTRIBUTED_CMD_ARGS_EXPECTED, distributed_command_with_args)
+
+
+if __name__ == "__main__":
+    from pyspark.ml.deepspeed.tests.test_deepspeed_distributor import *  # noqa: F401,F403
+
+    try:
+        import xmlrunner  # type:ignore
+
+        testRunner = xmlrunner.XMLTestRunner(output="target/test-reports", verbosity=2)
+    except ImportError:
+        testRunner = None
+    unittest.main(testRunner=testRunner, verbosity=2)