Deprecate log_gpu_memory, gpu_metrics, and util funcs in favor of DeviceStatsMonitor callback (#9921)

Co-authored-by: Kaushik B <[email protected]>
Co-authored-by: Rohit Gupta <[email protected]>

Author: 3 people

CHANGELOG.md

@@ -345,7 +345,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Deprecated passing `progress_bar_refresh_rate` to the `Trainer` constructor in favor of adding the `ProgressBar` callback with `refresh_rate` directly to the list of callbacks, or passing `enable_progress_bar=False` to disable the progress bar ([#9616](https://github.com/PyTorchLightning/pytorch-lightning/pull/9616))
 
 
-- Deprecate `LightningDistributed` and move the broadcast logic to `DDPPlugin` and `DDPSpawnPlugin` directly ([#9691](https://github.com/PyTorchLightning/pytorch-lightning/pull/9691))
+- Deprecated `LightningDistributed` and move the broadcast logic to `DDPPlugin` and `DDPSpawnPlugin` directly ([#9691](https://github.com/PyTorchLightning/pytorch-lightning/pull/9691))
 
 
 - Deprecated passing `stochastic_weight_avg` from the `Trainer` constructor in favor of adding the `StochasticWeightAveraging` callback directly to the list of callbacks ([#8989](https://github.com/PyTorchLightning/pytorch-lightning/pull/8989))
@@ -366,6 +366,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Deprecated passing `weights_summary` to the `Trainer` constructor in favor of adding the `ModelSummary` callback with `max_depth` directly to the list of callbacks ([#9699](https://github.com/PyTorchLightning/pytorch-lightning/pull/9699))
 
 
+- Deprecated `log_gpu_memory`, `gpu_metrics`, and util funcs in favor of `DeviceStatsMonitor` callback ([#9921](https://github.com/PyTorchLightning/pytorch-lightning/pull/9921))
+
+
 - Deprecated `GPUStatsMonitor` and `XLAStatsMonitor` in favor of `DeviceStatsMonitor` callback ([#9924](https://github.com/PyTorchLightning/pytorch-lightning/pull/9924))
 
 ### Removed

docs/source/common/trainer.rst

@@ -528,7 +528,7 @@ Example::
 checkpoint_callback
 ^^^^^^^^^^^^^^^^^^^
 
-Deprecated: This has been deprecated in v1.5 and will be removed in v.17. Please use ``enable_checkpointing`` instead.
+Deprecated: This has been deprecated in v1.5 and will be removed in v1.7. Please use ``enable_checkpointing`` instead.
 
 default_root_dir
 ^^^^^^^^^^^^^^^^
@@ -838,36 +838,6 @@ How often to add logging rows (does not write to disk)
 See Also:
     - :doc:`logging <../extensions/logging>`
 
-log_gpu_memory
-^^^^^^^^^^^^^^
-
-.. raw:: html
-
-    <video width="50%" max-width="400px" controls
-    poster="https://pl-bolts-doc-images.s3.us-east-2.amazonaws.com/pl_docs/trainer_flags/thumb/log_gpu_memory.jpg"
-    src="https://pl-bolts-doc-images.s3.us-east-2.amazonaws.com/pl_docs/trainer_flags/log_gpu_memory.mp4"></video>
-
-|
-
-Options:
-
-- None
-- 'min_max'
-- 'all'
-
-.. testcode::
-
-    # default used by the Trainer
-    trainer = Trainer(log_gpu_memory=None)
-
-    # log all the GPUs (on master node only)
-    trainer = Trainer(log_gpu_memory="all")
-
-    # log only the min and max memory on the master node
-    trainer = Trainer(log_gpu_memory="min_max")
-
-.. note:: Might slow performance because it uses the output of ``nvidia-smi``.
-
 logger
 ^^^^^^
 

pytorch_lightning/trainer/connectors/logger_connector/logger_connector.py

@@ -29,6 +29,11 @@
 class LoggerConnector:
     def __init__(self, trainer: "pl.Trainer", log_gpu_memory: Optional[str] = None) -> None:
         self.trainer = trainer
+        if log_gpu_memory is not None:
+            rank_zero_deprecation(
+                "Setting `log_gpu_memory` with the trainer flag is deprecated in v1.5 and will be removed in v1.7. "
+                "Please monitor GPU stats with the `DeviceStatsMonitor` callback directly instead."
+            )
         self.log_gpu_memory = log_gpu_memory
         self.eval_loop_results: List[_OUT_DICT] = []
         self._val_log_step: int = 0
@@ -222,6 +227,7 @@ def update_train_step_metrics(self) -> None:
         if self.trainer.fit_loop._should_accumulate() and self.trainer.lightning_module.automatic_optimization:
             return
 
+        # TODO: remove this call in v1.7
         self._log_gpus_metrics()
 
         # when metrics should be logged
@@ -239,6 +245,11 @@ def update_train_epoch_metrics(self) -> None:
         self.trainer._results.reset(metrics=True)
 
     def _log_gpus_metrics(self) -> None:
+        """
+        .. deprecated:: v1.5
+            This function was deprecated in v1.5 in favor of
+            `pytorch_lightning.accelerators.gpu._get_nvidia_gpu_stats` and will be removed in v1.7.
+        """
         for key, mem in self.gpus_metrics.items():
             if self.log_gpu_memory == "min_max":
                 self.trainer.lightning_module.log(key, mem, prog_bar=False, logger=True)
@@ -309,6 +320,14 @@ def metrics(self) -> _METRICS:
 
     @property
     def gpus_metrics(self) -> Dict[str, float]:
+        """
+        .. deprecated:: v1.5
+            Will be removed in v1.7.
+        """
+        rank_zero_deprecation(
+            "The property `LoggerConnector.gpus_metrics` was deprecated in v1.5"
+            " and will be removed in 1.7. Use the `DeviceStatsMonitor` callback instead."
+        )
         if self.trainer._device_type == DeviceType.GPU and self.log_gpu_memory:
             mem_map = memory.get_memory_profile(self.log_gpu_memory)
             self._gpus_metrics.update(mem_map)

pytorch_lightning/trainer/trainer.py

@@ -134,7 +134,7 @@ def __init__(
         auto_select_gpus: bool = False,
         tpu_cores: Optional[Union[List[int], str, int]] = None,
         ipus: Optional[int] = None,
-        log_gpu_memory: Optional[str] = None,
+        log_gpu_memory: Optional[str] = None,  # TODO: Remove in 1.7
         progress_bar_refresh_rate: Optional[int] = None,  # TODO: remove in v1.7
         enable_progress_bar: bool = True,
         overfit_batches: Union[int, float] = 0.0,
@@ -277,6 +277,10 @@ def __init__(
 
             log_gpu_memory: None, 'min_max', 'all'. Might slow performance.
 
+                .. deprecated:: v1.5
+                    Deprecated in v1.5.0 and will be removed in v1.7.0
+                    Please use the ``DeviceStatsMonitor`` callback directly instead.
+
             log_every_n_steps: How often to log within steps (defaults to every 50 steps).
 
             prepare_data_per_node: If True, each LOCAL_RANK=0 will call prepare data.

pytorch_lightning/utilities/memory.py

@@ -96,7 +96,12 @@ def garbage_collection_cuda() -> None:
 
 
 def get_memory_profile(mode: str) -> Dict[str, float]:
-    """Get a profile of the current memory usage.
+    r"""
+    .. deprecated:: v1.5
+        This function was deprecated in v1.5 in favor of
+        `pytorch_lightning.accelerators.gpu._get_nvidia_gpu_stats` and will be removed in v1.7.
+
+    Get a profile of the current memory usage.
 
     Args:
         mode: There are two modes:
@@ -124,7 +129,12 @@ def get_memory_profile(mode: str) -> Dict[str, float]:
 
 
 def get_gpu_memory_map() -> Dict[str, float]:
-    """Get the current gpu usage.
+    r"""
+    .. deprecated:: v1.5
+        This function was deprecated in v1.5 in favor of
+        `pytorch_lightning.accelerators.gpu._get_nvidia_gpu_stats` and will be removed in v1.7.
+
+    Get the current gpu usage.
 
     Return:
         A dictionary in which the keys are device ids as integers and

tests/deprecated_api/test_remove_1-7.py

@@ -21,6 +21,7 @@
 from pytorch_lightning.callbacks.gpu_stats_monitor import GPUStatsMonitor
 from pytorch_lightning.callbacks.xla_stats_monitor import XLAStatsMonitor
 from pytorch_lightning.loggers import LoggerCollection, TestTubeLogger
+from pytorch_lightning.trainer.connectors.logger_connector import LoggerConnector
 from tests.deprecated_api import _soft_unimport_module
 from tests.helpers import BoringModel
 from tests.helpers.datamodules import MNISTDataModule
@@ -370,6 +371,16 @@ def test_v1_7_0_weights_summary_trainer(tmpdir):
         t.weights_summary = "blah"
 
 
+def test_v1_7_0_trainer_log_gpu_memory(tmpdir):
+    with pytest.deprecated_call(
+        match="Setting `log_gpu_memory` with the trainer flag is deprecated in v1.5 and will be removed"
+    ):
+        trainer = Trainer(log_gpu_memory="min_max")
+    with pytest.deprecated_call(match="The property `LoggerConnector.gpus_metrics` was deprecated in v1.5"):
+        lg = LoggerConnector(trainer)
+        _ = lg.gpus_metrics
+
+
 @RunIf(min_gpus=1)
 def test_v1_7_0_deprecate_gpu_stats_monitor(tmpdir):
     with pytest.deprecated_call(match="The `GPUStatsMonitor` callback was deprecated in v1.5"):