Fix typos: deepseed -> deepspeed (#2499)

Author: jinyouzhi

deepspeed/runtime/pipe/module.py

@@ -119,7 +119,7 @@ def forward(self, inputs):
         Args:
             layers (Iterable): A sequence of layers defining pipeline structure. Can be a ``torch.nn.Sequential`` module.
             num_stages (int, optional): The degree of pipeline parallelism. If not specified, ``topology`` must be provided.
-            topology (``deepseed.runtime.pipe.ProcessTopology``, optional): Defines the axes of parallelism axes for training. Must be provided if ``num_stages`` is ``None``.
+            topology (``deepspeed.runtime.pipe.ProcessTopology``, optional): Defines the axes of parallelism axes for training. Must be provided if ``num_stages`` is ``None``.
             loss_fn (callable, optional): Loss is computed ``loss = loss_fn(outputs, label)``
             base_seed (int, optional): [description]. Defaults to 1234.
             partition_method (str, optional): [description]. Defaults to 'parameters'.

deepspeed/runtime/zero/partition_parameters.py

@@ -605,7 +605,7 @@ def __init__(self,
 
         .. note::
             Initializes ``deepspeed.comm`` if it has not already been done so.
-            See :meth:`deepseed.init_distributed` for more information.
+            See :meth:`deepspeed.init_distributed` for more information.
 
         .. note::
             Can also be used as a decorator:

docs/_tutorials/large-models-w-deepspeed.md

@@ -24,7 +24,7 @@ At a broad level, there are two primary paths to training a large model:
 
 Since, ZeRO is a replacement to data parallelism, it offers a seamless integration that does not require model code refactoring for existing data-parallel models. For majority of cases, ZeRO based technologies offers model scalability, training throughput efficiency without compromising ease of use.
 
-**3D Parallelism based technologies**: 3D Parallelism refers to a combination of three different forms of parallel technologies namely tensor-slicing, pipeline-parallelism, and data parallelism (or ZeRO powered data parallelism). Combing these three forms allows for harnessing the strength of each of these technologies without the drawback of any. 3D Parallelism enables DeepSeed to achieve excellent training throughput efficiency in the scenarios where relying on ZeRO based technologies alone might be insufficient. However, 3D parallelism requires non-trivial model code refactoring, and therefore a careful consideration is important to identify cases where 3D-Parallelism can bring non-trivial throughput benefits.
+**3D Parallelism based technologies**: 3D Parallelism refers to a combination of three different forms of parallel technologies namely tensor-slicing, pipeline-parallelism, and data parallelism (or ZeRO powered data parallelism). Combing these three forms allows for harnessing the strength of each of these technologies without the drawback of any. 3D Parallelism enables DeepSpeed to achieve excellent training throughput efficiency in the scenarios where relying on ZeRO based technologies alone might be insufficient. However, 3D parallelism requires non-trivial model code refactoring, and therefore a careful consideration is important to identify cases where 3D-Parallelism can bring non-trivial throughput benefits.
 
 ## Deciding which technology to use