Merge branch 'master' into master

Author: timyber

doc/api/prep_data/feature_store.rst

@@ -73,6 +73,10 @@ Inputs
     :members:
     :show-inheritance:
 
+.. autoclass:: sagemaker.feature_store.inputs.TableFormatEnum
+    :members:
+    :show-inheritance:
+
 
 Dataset Builder
 ***************

doc/api/training/sdp_versions/latest.rst

@@ -26,8 +26,8 @@ depending on the version of the library you use.
    <https://docs.aws.amazon.com/sagemaker/latest/dg/data-parallel-use-api.html#data-parallel-use-python-skd-api>`_
    for more information.
 
-Version 1.4.0, 1.4.1, 1.5.0 (Latest)
-====================================
+Version 1.4.0, 1.4.1, 1.5.0, 1.6.0 (Latest)
+===========================================
 
 .. toctree::
    :maxdepth: 1

doc/api/training/smd_data_parallel_release_notes/smd_data_parallel_change_log.rst

@@ -7,9 +7,51 @@ Release Notes
 New features, bug fixes, and improvements are regularly made to the SageMaker
 distributed data parallel library.
 
-SageMaker Distributed Data Parallel 1.5.0 Release Notes
+SageMaker Distributed Data Parallel 1.6.0 Release Notes
 =======================================================
 
+*Date: Dec. 15. 2022*
+
+**New Features**
+
+* New optimized SMDDP AllGather collective to complement the sharded data parallelism technique
+  in the SageMaker model parallelism library. For more information, see `Sharded data parallelism with SMDDP Collectives
+  <https://docs.aws.amazon.com/sagemaker/latest/dg/model-parallel-extended-features-pytorch-sharded-data-parallelism.html#model-parallel-extended-features-pytorch-sharded-data-parallelism-smddp-collectives>`_
+  in the *Amazon SageMaker Developer Guide*.
+* Added support for Amazon EC2 ``ml.p4de.24xlarge`` instances. You can run data parallel training jobs
+  on ``ml.p4de.24xlarge`` instances with the SageMaker data parallelism library’s AllReduce collective.
+
+**Improvements**
+
+* General performance improvements of the SMDDP AllReduce collective communication operation.
+
+**Migration to AWS Deep Learning Containers**
+
+This version passed benchmark testing and is migrated to the following AWS Deep Learning Containers (DLC):
+
+- SageMaker training container for PyTorch v1.12.1
+
+  .. code::
+
+    763104351884.dkr.ecr.<region>.amazonaws.com/pytorch-training:1.12.1-gpu-py38-cu113-ubuntu20.04-sagemaker
+
+
+Binary file of this version of the library for `custom container
+<https://docs.aws.amazon.com/sagemaker/latest/dg/data-parallel-use-api.html#data-parallel-bring-your-own-container>`_ users:
+
+  .. code::
+
+    https://smdataparallel.s3.amazonaws.com/binary/pytorch/1.12.1/cu113/2022-12-05/smdistributed_dataparallel-1.6.0-cp38-cp38-linux_x86_64.whl
+
+
+----
+
+Release History
+===============
+
+SageMaker Distributed Data Parallel 1.5.0 Release Notes
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
 *Date: Jul. 26. 2022*
 
 **Currency Updates**
@@ -38,12 +80,6 @@ Binary file of this version of the library for `custom container
 
     https://smdataparallel.s3.amazonaws.com/binary/pytorch/1.12.0/cu113/2022-07-01/smdistributed_dataparallel-1.5.0-cp38-cp38-linux_x86_64.whl
 
-
-----
-
-Release History
-===============
-
 SageMaker Distributed Data Parallel 1.4.1 Release Notes
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 

doc/api/training/smd_model_parallel_release_notes/smd_model_parallel_change_log.rst

@@ -6,9 +6,60 @@ New features, bug fixes, and improvements are regularly made to the SageMaker
 distributed model parallel library.
 
 
-SageMaker Distributed Model Parallel 1.11.0 Release Notes
+SageMaker Distributed Model Parallel 1.13.0 Release Notes
 =========================================================
 
+*Date: Dec. 15. 2022*
+
+**New Features**
+
+* Sharded data parallelism now supports a new backend for collectives called *SMDDP Collectives*.
+  For supported scenarios, SMDDP Collectives are on by default for the AllGather operation.
+  For more information, see
+  `Sharded data parallelism with SMDDP Collectives
+  <https://docs.aws.amazon.com/sagemaker/latest/dg/model-parallel-extended-features-pytorch-sharded-data-parallelism.html#model-parallel-extended-features-pytorch-sharded-data-parallelism-smddp-collectives>`_
+  in the *Amazon SageMaker Developer Guide*.
+* Introduced FlashAttention for DistributedTransformer to improve memory usage and computational
+  performance of models such as GPT2, GPTNeo, GPTJ, GPTNeoX, BERT, and RoBERTa.
+
+**Bug Fixes**
+
+* Fixed initialization of ``lm_head`` in DistributedTransformer to use a provided range
+  for initialization, when weights are not tied with the embeddings.
+
+**Improvements**
+
+* When a module has no parameters, we have introduced an optimization to execute
+  such a module on the same rank as its parent during pipeline parallelism.
+
+**Migration to AWS Deep Learning Containers**
+
+This version passed benchmark testing and is migrated to the following AWS Deep Learning Containers (DLC):
+
+- SageMaker training container for PyTorch v1.12.1
+
+  .. code::
+
+    763104351884.dkr.ecr.<region>.amazonaws.com/pytorch-training:1.12.1-gpu-py38-cu113-ubuntu20.04-sagemaker
+
+
+Binary file of this version of the library for `custom container
+<https://docs.aws.amazon.com/sagemaker/latest/dg/model-parallel-sm-sdk.html#model-parallel-bring-your-own-container>`_ users:
+
+- For PyTorch 1.12.0
+
+  .. code::
+
+    https://sagemaker-distributed-model-parallel.s3.us-west-2.amazonaws.com/pytorch-1.12.1/build-artifacts/2022-12-08-21-34/smdistributed_modelparallel-1.13.0-cp38-cp38-linux_x86_64.whl
+
+----
+
+Release History
+===============
+
+SageMaker Distributed Model Parallel 1.11.0 Release Notes
+---------------------------------------------------------
+
 *Date: August. 17. 2022*
 
 **New Features**
@@ -41,12 +92,7 @@ Binary file of this version of the library for `custom container
 
   .. code::
 
-    https://sagemaker-distributed-model-parallel.s3.us-west-2.amazonaws.com/pytorch-1.12.0/build-artifacts/2022-08-12-16-58/smdistributed_modelparallel-1.11.0-cp38-cp38-linux_x86_64.whl
-
-----
-
-Release History
-===============
+    https://sagemaker-distribu
 
 SageMaker Distributed Model Parallel 1.10.1 Release Notes
 ---------------------------------------------------------

doc/api/training/smp_versions/latest.rst

@@ -10,8 +10,8 @@ depending on which version of the library you need to use.
 To use the library, reference the
 **Common API** documentation alongside the framework specific API documentation.
 
-Version 1.11.0 (Latest)
-===========================================
+Version 1.11.0, 1.13.0 (Latest)
+===============================
 
 To use the library, reference the Common API documentation alongside the framework specific API documentation.
 

src/sagemaker/debugger/profiler_config.py

@@ -32,6 +32,7 @@ def __init__(
         s3_output_path: Optional[Union[str, PipelineVariable]] = None,
         system_monitor_interval_millis: Optional[Union[int, PipelineVariable]] = None,
         framework_profile_params: Optional[FrameworkProfile] = None,
+        disable_profiler: Optional[Union[str, PipelineVariable]] = False,
     ):
         """Initialize a ``ProfilerConfig`` instance.
 
@@ -78,6 +79,7 @@ class and SageMaker Framework estimators.
         self.s3_output_path = s3_output_path
         self.system_monitor_interval_millis = system_monitor_interval_millis
         self.framework_profile_params = framework_profile_params
+        self.disable_profiler = disable_profiler
 
     def _to_request_dict(self):
         """Generate a request dictionary using the parameters provided when initializing the object.
@@ -91,6 +93,8 @@ def _to_request_dict(self):
         if self.s3_output_path is not None:
             profiler_config_request["S3OutputPath"] = self.s3_output_path
 
+        profiler_config_request["DisableProfiler"] = self.disable_profiler
+
         if self.system_monitor_interval_millis is not None:
             profiler_config_request[
                 "ProfilingIntervalInMilliseconds"

src/sagemaker/estimator.py

@@ -938,26 +938,29 @@ def _prepare_collection_configs(self):
     def _prepare_profiler_for_training(self):
         """Set necessary values and do basic validations in profiler config and profiler rules.
 
-        When user explicitly set rules to an empty list, default profiler rule won't be enabled.
-        Default profiler rule will be enabled in supported regions when either:
-        1. user doesn't specify any rules, i.e., rules=None; or
-        2. user only specify debugger rules, i.e., rules=[Rule.sagemaker(...)]
+        No default profiler rule will be used. The user needs to specify rules explicitly
         """
         if self.disable_profiler:
-            if self.profiler_config:
-                raise RuntimeError("profiler_config cannot be set when disable_profiler is True.")
+            if self.profiler_config and not self.profiler_config.disable_profiler:
+                raise RuntimeError(
+                    "profiler_config.disable_profiler cannot be False"
+                    + " when disable_profiler is True."
+                )
             if self.profiler_rules:
                 raise RuntimeError("ProfilerRule cannot be set when disable_profiler is True.")
         elif _region_supports_profiler(self.sagemaker_session.boto_region_name):
             if self.profiler_config is None:
                 self.profiler_config = ProfilerConfig(s3_output_path=self.output_path)
             if self.rules is None or (self.rules and not self.profiler_rules):
-                self.profiler_rules = [get_default_profiler_rule()]
+                self.profiler_rules = []
 
         if self.profiler_config and not self.profiler_config.s3_output_path:
             self.profiler_config.s3_output_path = self.output_path
 
         self.profiler_rule_configs = self._prepare_profiler_rules()
+        # if profiler_config is still None, it means the job has profiler disabled
+        if self.profiler_config is None:
+            self.profiler_config = ProfilerConfig(disable_profiler=True)
 
     def _prepare_profiler_rules(self):
         """Set any necessary values in profiler rules, if they are provided."""
@@ -1048,7 +1051,7 @@ def latest_job_profiler_artifacts_path(self):
             error_message="""Cannot get the profiling output artifacts path.
         The Estimator is not associated with a training job."""
         )
-        if self.profiler_config is not None:
+        if self.profiler_config is not None and not self.profiler_config.disable_profiler:
             return os.path.join(
                 self.profiler_config.s3_output_path,
                 self.latest_training_job.name,
@@ -1895,8 +1898,8 @@ def enable_default_profiling(self):
         else:
             self.profiler_config = ProfilerConfig(s3_output_path=self.output_path)
 
-        self.profiler_rules = [get_default_profiler_rule()]
-        self.profiler_rule_configs = self._prepare_profiler_rules()
+        self.profiler_rules = []
+        self.profiler_rule_configs = []
 
         _TrainingJob.update(
             self, self.profiler_rule_configs, self.profiler_config._to_request_dict()

src/sagemaker/fw_utils.py

@@ -80,6 +80,7 @@
     "ml.p3.16xlarge",
     "ml.p3dn.24xlarge",
     "ml.p4d.24xlarge",
+    "ml.p4de.24xlarge",
     "local_gpu",
 )
 SM_DATAPARALLEL_SUPPORTED_FRAMEWORK_VERSIONS = {

src/sagemaker/session.py

@@ -3336,6 +3336,11 @@ def create_endpoint_config_from_existing(
         if request_data_capture_config_dict is not None:
             request["DataCaptureConfig"] = request_data_capture_config_dict
 
+        if existing_endpoint_config_desc.get("AsyncInferenceConfig") is not None:
+            request["AsyncInferenceConfig"] = existing_endpoint_config_desc.get(
+                "AsyncInferenceConfig", None
+            )
+
         self.sagemaker_client.create_endpoint_config(**request)
 
     def create_endpoint(self, endpoint_name, config_name, tags=None, wait=True):

tests/integ/sagemaker/workflow/test_workflow.py

@@ -1269,8 +1269,6 @@ def test_caching_behavior(
         # create pipeline
         pipeline.create(role)
         definition = json.loads(pipeline.definition())
-        # delete profiler config for assertions as it will contain a timestamp
-        del definition["Steps"][1]["Arguments"]["ProfilerRuleConfigurations"]
 
         # verify input path
         expected_abalone_input_path = f"{pipeline_name}/{step_process.name}" f"/input/abalone_data"
@@ -1295,7 +1293,6 @@ def test_caching_behavior(
 
         # verify no changes
         definition2 = json.loads(pipeline.definition())
-        del definition2["Steps"][1]["Arguments"]["ProfilerRuleConfigurations"]
         assert definition == definition2
 
         # add dummy file to source_dir
@@ -1306,7 +1303,6 @@ def test_caching_behavior(
 
         # verify changes
         definition3 = json.loads(pipeline.definition())
-        del definition3["Steps"][1]["Arguments"]["ProfilerRuleConfigurations"]
         assert definition != definition3
 
     finally: