Fix typos in docstrings

core/custom_training_loop.py

@@ -115,7 +115,7 @@ def train(
     dataset: data iterator for the training set
     evaluation_iterators: data iterators for the different evaluation sets
     scheduler: optional learning rate scheduler
-    output_transform_for_metrics: optional transformation functions to transorm the model
+    output_transform_for_metrics: optional transformation functions to transform the model
                                   output and labels into a format the metrics can understand
   """
 

core/debug_training_loop.py

@@ -1,6 +1,6 @@
 """This is a very limited feature training loop useful for interactive debugging.
 
-It is not intended for actual model tranining (it is not fast, doesn't compile the model).
+It is not intended for actual model training (it is not fast, doesn't compile the model).
 It does not support checkpointing.
 
 suggested use:

core/train_pipeline.py

@@ -57,7 +57,7 @@ def _wait_for_batch(batch: In, stream: Optional[torch.cuda.streams.Stream]) -> N
   torch.cuda.current_stream().wait_stream(stream)
   # As mentioned in https://pytorch.org/docs/stable/generated/torch.Tensor.record_stream.html,
   # PyTorch uses the "caching allocator" for memory allocation for tensors. When a tensor is
-  # freed, its memory is likely to be reused by newly constructed tenosrs.  By default,
+  # freed, its memory is likely to be reused by newly constructed tensors.  By default,
   # this allocator traces whether a tensor is still in use by only the CUDA stream where it
   # was created.   When a tensor is used by additional CUDA streams, we need to call record_stream
   # to tell the allocator about all these streams.  Otherwise, the allocator might free the