maddiedawson commented on code in PR #41770:
URL: https://github.com/apache/spark/pull/41770#discussion_r1258700232
##########
python/pyspark/ml/torch/tests/test_distributor.py:
##########
@@ -543,7 +543,6 @@ def test_check_parent_alive(self, mock_clean_and_terminate:
Callable) -> None:
class TorchWrapperUnitTests(TorchWrapperUnitTestsMixin, unittest.TestCase):
pass
-
Review Comment:
Undo this change
##########
python/pyspark/ml/torch/distributor.py:
##########
@@ -39,6 +39,9 @@
Generator,
Iterator,
)
+import shutil
+import subprocess
Review Comment:
These are already imported
##########
python/pyspark/ml/torch/deepspeed/tests/test_deepspeed_distributor.py:
##########
@@ -0,0 +1,145 @@
+import os
+import sys
+from typing import List, Any, Callable, Dict
+import unittest
+
+from pyspark.ml.torch.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)
+ value = default_value
+ return value
+
+ def _get_env_variables_distributed(self):
+ 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):
+ 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):
+ 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):
+ DEEPSPEED_CONF = "path/to/deepspeed"
+ TRAIN_FILE_PATH = "path/to/exec"
+ NUM_PROCS = 10
+ input_params = {}
+ input_params["local_mode"] = True
+ input_params['num_processes'] = NUM_PROCS
+ input_params["deepspeed_config"] = DEEPSPEED_CONF
+
+ # get the arguments for no argument, local run
+ torchrun_local_args_expected = ["--standalone", "--nnodes=1"]
+
+ 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)
+ 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)
Review Comment:
(Optional) Personally, I would split this into two test cases since there
isn't much shared setup or validation here. Same with the distributed test
below. The input arg creation could be deduped using a helper function
##########
python/pyspark/ml/torch/deepspeed/deepspeed_distributor.py:
##########
@@ -0,0 +1,144 @@
+
Review Comment:
Remove extra newline
##########
python/pyspark/ml/torch/deepspeed/deepspeed_distributor.py:
##########
@@ -0,0 +1,144 @@
+
+import json
+import os
+import sys
+import tempfile
+from typing import (
+ Union,
+ Callable,
+ List,
+ Dict,
+ Optional,
+ Any,
+ Tuple,
+)
+
+from pyspark.ml.torch.distributor import TorchDistributor
+
+class DeepspeedTorchDistributor(TorchDistributor):
+
+ def __init__(self, num_gpus: int = 1, nnodes: int = 1, local_mode: bool =
True, use_gpu: bool = True, deepspeed_config = None):
Review Comment:
Let's put each arg on its own line since there are a lot
##########
python/pyspark/ml/torch/deepspeed/deepspeed_distributor.py:
##########
@@ -0,0 +1,144 @@
+
+import json
+import os
+import sys
+import tempfile
+from typing import (
+ Union,
+ Callable,
+ List,
+ Dict,
+ Optional,
+ Any,
+ Tuple,
+)
+
+from pyspark.ml.torch.distributor import TorchDistributor
+
+class DeepspeedTorchDistributor(TorchDistributor):
+
+ def __init__(self, num_gpus: int = 1, nnodes: int = 1, local_mode: bool =
True, use_gpu: bool = True, deepspeed_config = 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 is a dictionary mapping parameters to values, then we
will create the file.
+ If None, deepspeed will fall back to default parameters.
+ """
+ num_processes = num_gpus * nnodes
+ super().__init__(num_processes, local_mode, use_gpu)
+ self.deepspeed_config = deepspeed_config
+ self.ssl_conf = "deepspeed.spark.distributor.ignoreSsl"
+ self._validate_input_params()
+ self.input_params = self._create_input_params()
+ self.cleanup_deepspeed_conf = False
+
+ @staticmethod
+ def _get_deepspeed_config_path(deepspeed_config):
+ if isinstance(deepspeed_config, dict):
+ with tempfile.NamedTemporaryFile(mode='w+', delete=False,
suffix='.json') as fil:
+ json.dump(deepspeed_config, fil)
+ return fil.name
+ deepspeed_config_path = deepspeed_config
+ # Empty value means the deepspeed will fall back to default settings.
+ if deepspeed_config == None:
+ return ""
+ return deepspeed_config_path
+
+
+ @staticmethod
+ def _get_torchrun_args(local_mode: bool, num_processes: int) ->
Tuple[List[Any], int]:
+ """
+ Given the mode and the number of processes, create the arguments to be
given to deepspeed
+
+ Parameters
+ ---------
+ local_mode: bool
+ Whether or not we are running training locally or in a distributed
fashion
+
+ num_processes: int
+ The number of processes that we are going to use (number of gpus
per node * number of nodes)
+
+ Returns
+ ------
+ Tuple[List[Any], int]
+ A tuple containing a list of arguments to pass as pytorch args to
deepspeed, as well as the number of processes per node
+ """
+ if local_mode:
+ torchrun_args = ["--standalone","--nnodes=1"]
+ processes_per_node = num_processes
+ return torchrun_args, processes_per_node
+
+ master_addr = os.environ["MASTER_ADDR"]
+ master_port = os.environ["MASTER_PORT"]
+ node_rank = os.environ["RANK"]
+ torchrun_args = [
+ f"--nnodes={num_processes}",
+ f"--node_rank={node_rank}",
+ f"--rdzv_endpoint={master_addr}:{master_port}",
+ "--rdzv_id=0",
+ ]
+ processes_per_node = 1
+ return torchrun_args, processes_per_node
Review Comment:
It looks like these are the same args created in TorchDistributor's
_create_torchrun_command. Let's move this function to that class and call it
from _create_torchrun_command
##########
python/pyspark/ml/torch/deepspeed/deepspeed_distributor.py:
##########
@@ -0,0 +1,144 @@
+
+import json
+import os
+import sys
+import tempfile
+from typing import (
+ Union,
+ Callable,
+ List,
+ Dict,
+ Optional,
+ Any,
+ Tuple,
+)
+
+from pyspark.ml.torch.distributor import TorchDistributor
+
+class DeepspeedTorchDistributor(TorchDistributor):
+
+ def __init__(self, num_gpus: int = 1, nnodes: int = 1, local_mode: bool =
True, use_gpu: bool = True, deepspeed_config = 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 is a dictionary mapping parameters to values, then we
will create the file.
+ If None, deepspeed will fall back to default parameters.
+ """
+ num_processes = num_gpus * nnodes
+ super().__init__(num_processes, local_mode, use_gpu)
+ self.deepspeed_config = deepspeed_config
+ self.ssl_conf = "deepspeed.spark.distributor.ignoreSsl"
+ self._validate_input_params()
+ self.input_params = self._create_input_params()
Review Comment:
We can remove these since they are in super's init function
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
