Avoid false positive warning about using sync_dist when using torch…

…metrics (#14143)

Co-authored-by: Carlos Mocholí <carlossmocholi@gmail.com>

src/pytorch_lightning/CHANGELOG.md

@@ -70,6 +70,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Fixed the device placement when `LightningModule.cuda()` gets called without specifying a device index and the current cuda device was not 0 ([#14128](https://github.com/Lightning-AI/lightning/pull/14128))
 
 
+- Avoided false positive warning about using `sync_dist` when using torchmetrics ([#14143](https://github.com/Lightning-AI/lightning/pull/14143))
+
+
 - Avoid `metadata.entry_points` deprecation warning on Python 3.10 ([#14052](https://github.com/Lightning-AI/lightning/pull/14052))
 
 
@@ -79,7 +82,6 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Fixed saving hyperparameters in a composition where the parent class is not a `LightningModule` or `LightningDataModule` ([#14151](https://github.com/Lightning-AI/lightning/pull/14151))
 
 
-
 ## [1.7.1] - 2022-08-09
 
 ### Fixed

src/pytorch_lightning/trainer/connectors/logger_connector/result.py

@@ -525,7 +525,7 @@ def _get_cache(result_metric: _ResultMetric, on_step: bool) -> Optional[Tensor]:
         elif not on_step and result_metric.meta.on_epoch:
             if result_metric._computed is None:
                 should = result_metric.meta.sync.should
-                if not result_metric.meta.sync.should and distributed_available():
+                if not should and distributed_available() and result_metric.is_tensor:
                     # ensure sync happens for FT since during a failure, the metrics are synced and saved to the
                     # checkpoint, so during restart, metrics on rank 0 are from the accumulated ones from the previous
                     # run, and on other ranks, they are 0. So we need to make sure they are synced in further training

tests/tests_pytorch/core/test_metric_result_integration.py

@@ -21,9 +21,11 @@
 import torch
 import torch.distributed as dist
 import torch.multiprocessing as mp
+import torchmetrics
 from torch.nn import ModuleDict, ModuleList
 from torchmetrics import Metric, MetricCollection
 
+import pytorch_lightning as pl
 import tests_pytorch.helpers.utils as tutils
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks import ModelCheckpoint
@@ -666,19 +668,27 @@ def on_train_start(self):
 
 
 @pytest.mark.parametrize("distributed_env", [True, False])
-def test_logger_sync_dist(distributed_env):
-    # self.log('bar', 7, ..., sync_dist=False)
+@pytest.mark.parametrize("log_val", [torch.tensor(0.5), torchmetrics.Accuracy()])
+def test_logger_sync_dist(distributed_env, log_val):
+    pl.trainer.connectors.logger_connector.result.warning_cache.clear()
+
+    # self.log('bar', 0.5, ..., sync_dist=False)
     meta = _Metadata("foo", "bar")
     meta.sync = _Sync(_should=False)
-    result_metric = _ResultMetric(metadata=meta, is_tensor=True)
-    result_metric.update(torch.tensor(7.0), 10)
+    is_tensor = isinstance(log_val, torch.Tensor)
+
+    if not is_tensor:
+        log_val.update(torch.tensor([0, 1]), torch.tensor([0, 0], dtype=torch.long))
+
+    result_metric = _ResultMetric(metadata=meta, is_tensor=is_tensor)
+    result_metric.update(log_val, 10)
 
-    warning_ctx = pytest.warns if distributed_env else no_warning_call
+    warning_ctx = pytest.warns if distributed_env and is_tensor else no_warning_call
 
     with mock.patch(
         "pytorch_lightning.trainer.connectors.logger_connector.result.distributed_available",
         return_value=distributed_env,
     ):
         with warning_ctx(PossibleUserWarning, match=r"recommended to use `self.log\('bar', ..., sync_dist=True\)`"):
             value = _ResultCollection._get_cache(result_metric, on_step=False)
-        assert value == 7.0
+        assert value == 0.5