auto commit: wip

docs/reference/experiment-config-reference.rst

@@ -182,19 +182,20 @@ Example:
 
 .. _scheduling-unit:
 
-``scheduling_unit``
-===================
+``scheduling_unit`` (deprecated)
+================================
 
 Optional. Instructs how frequent to perform system operations, such as periodic checkpointing and
-preemption, in the unit of batches. The number of records in a batch is controlled by the
-:ref:`global_batch_size <config-global-batch-size>` hyperparameter. Defaults to ``100``.
+preemption, in the unit of batches. This field has been deprecated and the behavior should be
+configured in training code directly. Please see :ref:`apis-howto-overview` for details specific to
+your training framework.
+
+.. _config-records-per-epoch:
 
--  Setting this value too small can increase the overhead of system operations and decrease training
-   throughput.
--  Setting this value too large might prevent the system from reallocating resources from this
-   workload to another, potentially more important, workload.
--  As a rule of thumb, it should be set to the number of batches that can be trained in roughly
-   60--180 seconds.
+``records_per_epoch`` (deprecated)
+==================================
+
+Optional. The number of records in the training data set. This field has been deprecated.
 
 .. _max-restarts:
 
@@ -319,6 +320,15 @@ While debugging, the logger will display lines highlighted in blue for easy iden
  Validation Policy
 *******************
 
+.. _experiment-config-min-validation-period:
+
+``min_validation_period`` (deprecated)
+======================================
+
+Optional. Specifies the minimum frequency at which validation should be run for each trial. This
+field has been deprecated and should be specified directly in training code. Please see
+:ref:`apis-howto-overview` for details specific to your training framework.
+
 .. _experiment-config-perform-initial-validation:
 
 ``perform_initial_validation``
@@ -341,6 +351,15 @@ Determined checkpoints in the following situations:
 -  Prior to the searcher making a decision based on the validation of trials, ensuring consistency
    in case of a failure.
 
+.. _experiment-config-min-checkpoint-period:
+
+``min_checkpoint_period`` (deprecated)
+======================================
+
+Optional. Specifies the minimum frequency for running checkpointing for each trial. This field has
+been deprecated and should be specified directly in training code. Please see
+:ref:`apis-howto-overview` for details specific to your training framework.
+
 ``checkpoint_policy``
 =====================
 

docs/release-notes/searcher-context-removal.rst

@@ -12,6 +12,10 @@
    and are now being removed. Users are encouraged to use a preset searcher, which can be easily
    :ref:`configured <experiment-configuration_searcher>` for any experiment.
 
+-  DeepSpeed: the ``num_micro_batches_per_slot`` and ``train_micro_batch_size_per_gpu`` attributes
+   on ``DeepSpeedContext`` have been replaced with ``get_train_micro_batch_size_per_gpu()`` and
+   ``get_num_micro_batches_per_slot()``.
+
 **New Features**
 
 -  API: introduce ``keras.DeterminedCallback``, a new high-level training API for TF Keras that
@@ -30,8 +34,9 @@
 -  Experiment Config: the ``optimizations`` config has been deprecated. Please see :ref:`Training
    APIs <apis-howto-overview>` to configure supported optimizations through training code directly.
 
--  Experiment Config: the ``scheduling_unit`` config field has been deprecated. min
-   checkpoint/val/records per epoch
+-  Experiment Config: the ``scheduling_unit``, ``min_checkpoint_period``, and
+   ``min_validation_period`` config fields have been deprecated. Instead, these configuration
+   options should be specified in training code.
 
 -  Experiment Config: the ``entrypoint`` field no longer accepts ``model_def:TrialClass`` as trial
    definitions. Please invoke your training script directly (``python3 train.py``).

harness/determined/pytorch/_trainer_utils.py

@@ -87,7 +87,7 @@ class Epoch(TrainUnit):
 
     Epoch(int) values are treated as periods, e.g. Epoch(100) will checkpoint/validate every 100
     epochs.
-    Epoch(collections.abc.Container) values are treated as schedules, e.g. Epoch(1,5,10) will
+    Epoch(collections.abc.Container) values are treated as schedules, e.g. Epoch([1,5,10]) will
     checkpoint/validate on epochs 1, 5, and 10.
     """
 
@@ -100,7 +100,7 @@ class Batch(TrainUnit):
 
     Batch(int) values are treated as periods, e.g. Batch(100) will checkpoint/validate every 100
     batches.
-    Batch(collections.abc.Container) values are treated as schedules, e.g. Batch(1,5,10) will
+    Batch(collections.abc.Container) values are treated as schedules, e.g. Batch([1,5,10]) will
     checkpoint/validate on batches 1, 5, and 10.
     """