fix: Fix minor comment/docstring typos in runtime and inference modules

Signed-off-by: nathon-lee <leejianwoo@gmail.com>

Author: nathon-lee

deepspeed/inference/v2/model_implementations/inference_model_base.py

@@ -233,7 +233,7 @@ def max_sequence_length(self) -> int:
 
     def maybe_free_kv(self, sequence: DSSequenceDescriptor) -> None:
         """
-        After completing a forward pass, determine whether or not the there are any KV blocks
+        After completing a forward pass, determine whether or not there are any KV blocks
         that maybe freed since they are no longer in use.
 
         Consider the following example:

deepspeed/runtime/activation_checkpointing/checkpointing.py

@@ -704,14 +704,14 @@ def backward(ctx, *grads):
 def non_reentrant_checkpoint(function, *args):
     """This function is union of `torch.utils.checkpoint._checkpoint_without_reentrant` and `CheckpointFunction` in this module
 
-    This function is aim to solve the back probagation error raised from all input requires no grad.
+    This function is aim to solve the back propagation error raised from all input requires no grad.
     * has already been implemented in pytorch for a while, the solution is stable at most time except for jit module mode.
     * can help to solve the issue which is hacked by `deepspeed.runtime.pipe.module.PipelineModule._is_checkpointable`
 
     Main modifications compared to the implementation of torch:
     1. adapt to the signature of `checkpoint` function in this module
     2. solve the non-deterministic by random state management consistent with deepspeed `CheckpointFunction`
-    3. when there is partition or cpu checkpointing, gather them in the unpack_hook during back probagation
+    3. when there is partition or cpu checkpointing, gather them in the unpack_hook during back propagation
     4. make all after backward blocks in the hook which will executed after all leaf nodes backward execution.
     5. above 4. is inspired by `torch.autograd.graph.register_multi_grad_hook`, which is only implemented after 2.0.0
     """
@@ -801,7 +801,7 @@ def checkpoint_unpack(holder_from_backward):
         """retrieve the activations from recompute"""
         nonlocal deepspeed_saved_tensors, non_tensor_args, tensor_flags
 
-        # if this is the first step of backward probagation, recompute the graph and save
+        # if this is the first step of backward propagation, recompute the graph and save
         # all the activations with the same order as `checkpoint_pack` does
         if len(storage) == 0:
             unpack_counter = 0

deepspeed/runtime/data_pipeline/data_sampling/variable_batch_size_and_lr.py

@@ -291,7 +291,7 @@ def step(self, epoch=None):
 
         self.base_lr_scheduler.step(epoch)  # set unscaled lr, _step_count, last_epoch, _last_lr for new epoch
 
-        # scale the learning rate for the the next iteration for each parameter group.
+        # scale the learning rate for the next iteration for each parameter group.
         self.last_epoch = self.last_epoch + 1 if epoch is None else epoch
         # batch sizes are precomputed and stored in batch_sizes se we loop around to get the next one
         batch_size = self.batch_sizes[self.last_epoch % len(self.batch_sizes)]