You are viewing a plain text version of this content. The canonical link for it is here.
Posted to reviews@spark.apache.org by GitBox <gi...@apache.org> on 2023/01/03 18:35:20 UTC
[GitHub] [spark] rithwik-db opened a new pull request, #39369: [WIP][SPARK-41775][PYTHON][ML] Adding support for PyForch functions
rithwik-db opened a new pull request, #39369:
URL: https://github.com/apache/spark/pull/39369
<!--
Thanks for sending a pull request! Here are some tips for you:
1. If this is your first time, please read our contributor guidelines: https://spark.apache.org/contributing.html
2. Ensure you have added or run the appropriate tests for your PR: https://spark.apache.org/developer-tools.html
3. If the PR is unfinished, add '[WIP]' in your PR title, e.g., '[WIP][SPARK-XXXX] Your PR title ...'.
4. Be sure to keep the PR description updated to reflect all changes.
5. Please write your PR title to summarize what this PR proposes.
6. If possible, provide a concise example to reproduce the issue for a faster review.
7. If you want to add a new configuration, please read the guideline first for naming configurations in
'core/src/main/scala/org/apache/spark/internal/config/ConfigEntry.scala'.
8. If you want to add or modify an error type or message, please read the guideline first in
'core/src/main/resources/error/README.md'.
-->
NOTE: If you want to only view the diff from the other WIP PR regarding the baseline changes, look at the LAST COMMIT in this PR's commit history (titled "WIP adding notebook functionality"). Since I am sending out parallel PRs that are related, you should view this commit to see the diff pertaining to this ticket.
### What changes were proposed in this pull request?
<!--
Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue.
If possible, please consider writing useful notes for better and faster reviews in your PR. See the examples below.
1. If you refactor some codes with changing classes, showing the class hierarchy will help reviewers.
2. If you fix some SQL features, you can provide some references of other DBMSes.
3. If there is design documentation, please add the link.
4. If there is a discussion in the mailing list, please add the link.
-->
This is an addition to https://github.com/apache/spark/pull/39299 to add support for using functions as the input for distributed training. The users would follow the first workflow in the [design document](https://docs.google.com/document/d/1QPO1Ly8WteL6aIPvVcR7Xne9qVtJiB3fdrRn7NwBcpA/edit#heading=h.8yvw9xq428fh) to run training.
### Why are the changes needed?
<!--
Please clarify why the changes are needed. For instance,
1. If you propose a new API, clarify the use case for a new API.
2. If you fix a bug, you can clarify why it is a bug.
-->
We want to make it easier for users to run distributed training from a notebook setting.
### Does this PR introduce _any_ user-facing change?
<!--
Note that it means *any* user-facing change including all aspects such as the documentation fix.
If yes, please clarify the previous behavior and the change this PR proposes - provide the console output, description and/or an example to show the behavior difference if possible.
If possible, please also clarify if this is a user-facing change compared to the released Spark versions or within the unreleased branches such as master.
If no, write 'No'.
-->
Users can now input a training function into the `PyTorchDistributor().run(...)` api. This will require a lot of additional documentation though since since we are internally using cloudpickle to run training so we need to be able to have the user's train() function be picklable.
### How was this patch tested?
<!--
If tests were added, say they were added here. Please make sure to add some test cases that check the changes thoroughly including negative and positive cases if possible.
If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future.
If tests were not added, please describe why they were not added and/or why it was difficult to add.
If benchmark tests were added, please run the benchmarks in GitHub Actions for the consistent environment, and the instructions could accord to: https://spark.apache.org/developer-tools.html#github-workflow-benchmarks.
-->
This is a WIP PR so it hasn't been tested yet.
--
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: reviews-unsubscribe@spark.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org
[GitHub] [spark] AmplabJenkins commented on pull request #39369: [WIP][SPARK-41775][PYTHON][ML] Adding support for PyForch functions
Posted by GitBox <gi...@apache.org>.
AmplabJenkins commented on PR #39369:
URL: https://github.com/apache/spark/pull/39369#issuecomment-1370472443
Can one of the admins verify this patch?
--
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: reviews-unsubscribe@spark.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org
[GitHub] [spark] rithwik-db commented on a diff in pull request #39369: [SPARK-41775][PYTHON][ML] Adding support for PyTorch functions
Posted by GitBox <gi...@apache.org>.
rithwik-db commented on code in PR #39369:
URL: https://github.com/apache/spark/pull/39369#discussion_r1082896210
##########
python/pyspark/ml/torch/tests/test_distributor.py:
##########
@@ -349,6 +434,13 @@ def test_get_num_tasks_distributed(self) -> None:
self.spark.sparkContext._conf.set("spark.task.resource.gpu.amount", "1")
+ # def test_end_to_end_run_distributedly(self) -> None:
Review Comment:
Sorry, that shouldn't have been commented out. Will quickly update adding that test back in
--
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: reviews-unsubscribe@spark.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org
[GitHub] [spark] dongjoon-hyun commented on a diff in pull request #39369: [SPARK-41775][PYTHON][ML] Adding support for PyTorch functions
Posted by "dongjoon-hyun (via GitHub)" <gi...@apache.org>.
dongjoon-hyun commented on code in PR #39369:
URL: https://github.com/apache/spark/pull/39369#discussion_r1084987239
##########
python/pyspark/ml/torch/distributor.py:
##########
@@ -15,16 +15,21 @@
# limitations under the License.
#
+import cloudpickle # type: ignore
Review Comment:
This should be our `cloudpickle`.
--
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: reviews-unsubscribe@spark.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org
[GitHub] [spark] HyukjinKwon commented on pull request #39369: [SPARK-41775][PYTHON][ML] Adding support for PyTorch functions
Posted by "HyukjinKwon (via GitHub)" <gi...@apache.org>.
HyukjinKwon commented on PR #39369:
URL: https://github.com/apache/spark/pull/39369#issuecomment-1399942532
Merged to master.
--
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: reviews-unsubscribe@spark.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org
[GitHub] [spark] WeichenXu123 commented on a diff in pull request #39369: [SPARK-41775][PYTHON][ML] Adding support for PyForch functions
Posted by GitBox <gi...@apache.org>.
WeichenXu123 commented on code in PR #39369:
URL: https://github.com/apache/spark/pull/39369#discussion_r1082443370
##########
python/pyspark/ml/torch/distributor.py:
##########
@@ -495,32 +546,119 @@ def set_gpus(context: "BarrierTaskContext") -> None:
def _run_distributed_training(
self,
- framework_wrapper_fn: Optional[Callable],
+ framework_wrapper_fn: Callable,
train_object: Union[Callable, str],
*args: Any,
) -> Optional[Any]:
if not framework_wrapper_fn:
raise RuntimeError("Unknown combination of parameters")
+
+ log_streaming_server = LogStreamingServer()
+ self.driver_address = get_driver_host(self.sc)
+ log_streaming_server.start()
+ time.sleep(1) # wait for the server to start
+ self.log_streaming_server_port = log_streaming_server.port
+
spark_task_function = self._get_spark_task_function(
framework_wrapper_fn, train_object, *args
)
self._check_encryption()
- result = (
- self.sc.parallelize(range(self.num_tasks), self.num_tasks)
- .barrier()
- .mapPartitions(spark_task_function)
- .collect()[0]
+ self.logger.info(
+ f"Started distributed training with {self.num_processes} executor proceses"
+ )
+ try:
+ result = (
+ self.sc.parallelize(range(self.num_tasks), self.num_tasks)
+ .barrier()
+ .mapPartitions(spark_task_function)
+ .collect()[0]
+ )
+ finally:
+ log_streaming_server.shutdown()
+ self.logger.info(
+ f"Finished distributed training with {self.num_processes} executor proceses"
)
return result
@staticmethod
def _run_training_on_pytorch_file(
input_params: Dict[str, Any], train_path: str, *args: Any
) -> None:
+ log_streaming_client = input_params.get("log_streaming_client", None)
training_command = TorchDistributor._create_torchrun_command(
input_params, train_path, *args
)
- TorchDistributor._execute_command(training_command)
+ TorchDistributor._execute_command(
+ training_command, log_streaming_client=log_streaming_client
+ )
+
+ @contextmanager
+ @staticmethod
+ def _setup_files(train_fn: Callable, *args: Any) -> Tuple[str, str]:
+ save_dir = TorchDistributor._create_save_dir()
+ pickle_file_path = TorchDistributor._save_pickled_function(save_dir, train_fn, *args)
+ output_file_path = os.path.join(save_dir, TorchDistributor.PICKLED_OUTPUT_FILE)
+ train_file_path = TorchDistributor._create_torchrun_train_file(
+ save_dir, pickle_file_path, output_file_path
+ )
+ try:
+ yield (train_file_path, output_file_path)
+ finally:
+ TorchDistributor._cleanup_files(save_dir)
+
+ @staticmethod
+ def _run_training_on_pytorch_function(
+ input_params: dict[str, Any], train_fn: Callable, *args: Any # TODO: change dict to Dict
+ ) -> Any:
+ with TorchDistributor._setup_files(train_fn, *args) as (train_file_path, output_file_path):
+ args = [] # type: ignore
+ TorchDistributor._run_training_on_pytorch_file(input_params, train_file_path, *args)
+ output = TorchDistributor._get_pickled_output(output_file_path)
+ return output
+
+ @staticmethod
+ def _create_save_dir() -> str:
+ # TODO: need to do this in a safe way to avoid issues during concurrent runs
+ return tempfile.mkdtemp()
+
+ @staticmethod
+ def _cleanup_files(save_dir: str) -> None:
+ shutil.rmtree(save_dir)
Review Comment:
pls add argument `ignore_errors=True`
--
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: reviews-unsubscribe@spark.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org
[GitHub] [spark] WeichenXu123 commented on a diff in pull request #39369: [WIP][SPARK-41775][PYTHON][ML] Adding support for PyForch functions
Posted by GitBox <gi...@apache.org>.
WeichenXu123 commented on code in PR #39369:
URL: https://github.com/apache/spark/pull/39369#discussion_r1080791404
##########
python/pyspark/ml/torch/distributor.py:
##########
@@ -574,6 +582,68 @@ def _run_training_on_pytorch_file(
training_command, log_streaming_client=log_streaming_client
)
+ @staticmethod
+ def _run_training_on_pytorch_function(
+ input_params: dict[str, Any], train_fn: Callable, *args: Any
+ ) -> Any:
+ save_dir = TorchDistributor._create_save_dir()
+ pickle_file_path = TorchDistributor._save_pickled_function(save_dir, train_fn, *args)
+ output_file_path = os.path.join(save_dir, TorchDistributor.PICKLED_OUTPUT_FILE)
+ train_file_path = TorchDistributor._create_torchrun_train_file(
+ save_dir, pickle_file_path, output_file_path
+ )
+ args = []
+
+ TorchDistributor._run_training_on_pytorch_file(input_params, train_file_path, *args)
+
+ output = TorchDistributor._get_pickled_output(output_file_path)
+ TorchDistributor._cleanup_files(save_dir)
+ return output
+
+ @staticmethod
+ def _create_save_dir() -> str:
+ # TODO: need to do this in a safe way to avoid issues during concurrent runs
+ return tempfile.mkdtemp()
+
+ @staticmethod
+ def _cleanup_files(save_dir: str) -> None:
Review Comment:
The method is not called when training routine raises error.
--
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: reviews-unsubscribe@spark.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org
[GitHub] [spark] lu-wang-dl commented on a diff in pull request #39369: [SPARK-41775][PYTHON][ML] Adding support for PyTorch functions
Posted by GitBox <gi...@apache.org>.
lu-wang-dl commented on code in PR #39369:
URL: https://github.com/apache/spark/pull/39369#discussion_r1082894756
##########
python/pyspark/ml/torch/tests/test_distributor.py:
##########
@@ -349,6 +434,13 @@ def test_get_num_tasks_distributed(self) -> None:
self.spark.sparkContext._conf.set("spark.task.resource.gpu.amount", "1")
+ # def test_end_to_end_run_distributedly(self) -> None:
Review Comment:
Clean up this?
--
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: reviews-unsubscribe@spark.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org
[GitHub] [spark] WeichenXu123 commented on a diff in pull request #39369: [SPARK-41775][PYTHON][ML] Adding support for PyForch functions
Posted by GitBox <gi...@apache.org>.
WeichenXu123 commented on code in PR #39369:
URL: https://github.com/apache/spark/pull/39369#discussion_r1082450887
##########
python/pyspark/ml/torch/tests/test_distributor.py:
##########
@@ -224,8 +293,10 @@ def setUp(self) -> None:
self.sc = SparkContext("local-cluster[2,2,1024]", class_name, conf=conf)
self.spark = SparkSession(self.sc)
+ self.mnist_dir_path = tempfile.mkdtemp()
def tearDown(self) -> None:
+ shutil.rmtree(self.mnist_dir_path)
os.unlink(self.tempFile.name)
Review Comment:
Nit: refine the variable name: `tempFile` --> `gpu_discovery_script_file`
--
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: reviews-unsubscribe@spark.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org
[GitHub] [spark] WeichenXu123 commented on a diff in pull request #39369: [WIP][SPARK-41775][PYTHON][ML] Adding support for PyForch functions
Posted by GitBox <gi...@apache.org>.
WeichenXu123 commented on code in PR #39369:
URL: https://github.com/apache/spark/pull/39369#discussion_r1080790917
##########
python/pyspark/ml/torch/distributor.py:
##########
@@ -574,6 +582,68 @@ def _run_training_on_pytorch_file(
training_command, log_streaming_client=log_streaming_client
)
+ @staticmethod
+ def _run_training_on_pytorch_function(
+ input_params: dict[str, Any], train_fn: Callable, *args: Any
+ ) -> Any:
+ save_dir = TorchDistributor._create_save_dir()
+ pickle_file_path = TorchDistributor._save_pickled_function(save_dir, train_fn, *args)
+ output_file_path = os.path.join(save_dir, TorchDistributor.PICKLED_OUTPUT_FILE)
+ train_file_path = TorchDistributor._create_torchrun_train_file(
+ save_dir, pickle_file_path, output_file_path
+ )
+ args = []
+
+ TorchDistributor._run_training_on_pytorch_file(input_params, train_file_path, *args)
+
+ output = TorchDistributor._get_pickled_output(output_file_path)
+ TorchDistributor._cleanup_files(save_dir)
+ return output
+
+ @staticmethod
+ def _create_save_dir() -> str:
+ # TODO: need to do this in a safe way to avoid issues during concurrent runs
+ return tempfile.mkdtemp()
Review Comment:
pls delete it when training completes or raises error.
you can convert it to a python context manager
--
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: reviews-unsubscribe@spark.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org
[GitHub] [spark] WeichenXu123 commented on pull request #39369: [WIP][SPARK-41775][PYTHON][ML] Adding support for PyForch functions
Posted by GitBox <gi...@apache.org>.
WeichenXu123 commented on PR #39369:
URL: https://github.com/apache/spark/pull/39369#issuecomment-1396422850
overall good.
--
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: reviews-unsubscribe@spark.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org
[GitHub] [spark] HyukjinKwon closed pull request #39369: [SPARK-41775][PYTHON][ML] Adding support for PyTorch functions
Posted by "HyukjinKwon (via GitHub)" <gi...@apache.org>.
HyukjinKwon closed pull request #39369: [SPARK-41775][PYTHON][ML] Adding support for PyTorch functions
URL: https://github.com/apache/spark/pull/39369
--
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: reviews-unsubscribe@spark.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org
[GitHub] [spark] HyukjinKwon commented on a diff in pull request #39369: [SPARK-41775][PYTHON][ML] Adding support for PyTorch functions
Posted by "HyukjinKwon (via GitHub)" <gi...@apache.org>.
HyukjinKwon commented on code in PR #39369:
URL: https://github.com/apache/spark/pull/39369#discussion_r1085070034
##########
python/pyspark/ml/torch/distributor.py:
##########
@@ -15,16 +15,21 @@
# limitations under the License.
#
+import cloudpickle # type: ignore
Review Comment:
+1 thanks!
--
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: reviews-unsubscribe@spark.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org