docs: minor spelling tweaks (microsoft#858)

deepspeed/profiling/flops_profiler/profiler.py

@@ -265,7 +265,7 @@ def del_extra_repr(module):
                 "Each module profile is listed after its name in the following order: \nnumber of parameters, percentage of total parameters, number of multiply-accumulate operations (MACs), percentage of total MACs, latency, percentage of total latency, number of floating point operations per second (FLOPS, computed as 2 * MACs / latency)."
             )
             print(
-                "Note: \n1. A module can have torch.nn.functional (e.g. to compute logits) along with submodules, thus making the difference between the parent's MACs(or latency) and the sum of its submodules'.\n2. Number of floating point operations is a theoretical estimation, thus FLOPS computed using that could be larger than the maximum system throught.\n"
+                "Note: \n1. A module can have torch.nn.functional (e.g. to compute logits) along with submodules, thus making the difference between the parent's MACs(or latency) and the sum of its submodules'.\n2. Number of floating point operations is a theoretical estimation, thus FLOPS computed using that could be larger than the maximum system throughput.\n"
             )
             print(self.model)
 

docs/_pages/features.md

@@ -37,7 +37,7 @@ and communication- efficient training. DeepSpeed supports a hybrid
 combination of data, model, and pipeline parallelism and has scaled to over
 [one trillion parameters using 3D parallelism]({{ site.press_release_v3 }}).
 Pipeline parallelism can also improve communication efficiency and has
-accelerated training by up to 7x on low-banwdith clusters.
+accelerated training by up to 7x on low-bandwidth clusters.
 
 
 ## Model Parallelism
@@ -256,9 +256,9 @@ This can be enabled by setting the following in the `deepspeed_config` file.
 
 ```
 
-###  Timing Activiation Checkpoint Functions
+###  Timing Activation Checkpoint Functions
 
-When activiation checkpoingint is enabled, profiling the forward and backward time of each checkpoint function can be enabled in the `deepspeed_config` file.
+When activation checkpointing is enabled, profiling the forward and backward time of each checkpoint function can be enabled in the `deepspeed_config` file.
 
 ```json
 {

docs/_tutorials/flops-profiler.md

@@ -37,11 +37,11 @@ Top 3 modules in params at depth 2 are {'Conv2d': '50.69 k', 'Linear': '11.01 k'
 Top 3 modules in latency at depth 2 are {'Conv2d': '11.37 ms', 'Linear': '5.27 ms', 'AvgPool2d': '5.02 ms'}
 
 ------------------------------ Detailed Profile ------------------------------
-Each module profile is listed after its name in the follwing order:
+Each module profile is listed after its name in the following order:
 number of parameters, percentage of total parameters, number of multiply-accumulate operations (MACs), percentage of total MACs, latency, percentage of total latency, number of floating point operations per second (FLOPS, computed as 2 * MACs / latency).
 Note:
 1. A module can have torch.nn.functional (e.g. to compute logits) along with submodules, thus making the difference between the parent's MACs(or latency) and the sum of its submodules'.
-2. Number of floating point operations is a theoretical estimation, thus FLOPS computed using that could be larger than the maximum system throught.
+2. Number of floating point operations is a theoretical estimation, thus FLOPS computed using that could be larger than the maximum system throughput.
 
 LeNet5(
   61.71 k, 100.00% Params, 439.56 MMACs, 100.00% MACs, 25.7 ms, 100.00% latency, 34.2 GFLOPS,
@@ -92,7 +92,7 @@ The DeepSpeed flops profiler can be used with the DeepSpeed runtime or as a stan
 
 ### Usage With the DeepSpeed Runtime
 
-When using DeepSpeed for model training, the flops profiler can be configured in the `deepspeed_config` file. No explict API calls are needed to use the profiler. Refer to [flops profiler](https://www.deepspeed.ai/docs/config-json/#flops-profiler) for details.
+When using DeepSpeed for model training, the flops profiler can be configured in the `deepspeed_config` file. No explicit API calls are needed to use the profiler. Refer to [flops profiler](https://www.deepspeed.ai/docs/config-json/#flops-profiler) for details.
 
 
 #### Example: Megatron-LM
@@ -131,11 +131,11 @@ Top 3 modules in params at depth 8 are {'ColumnParallelLinear': '7.35 M', 'RowPa
 Top 3 modules in latency at depth 8 are {'ColumnParallelLinear': '659.23 us', 'RowParallelLinear': '587.94 us', 'FusedScaleMaskSoftmax': '370.98 us'}
 
 ------------------------------ Detailed Profile ------------------------------
-Each module profile is listed after its name in the follwing order:
+Each module profile is listed after its name in the following order:
 number of parameters, percentage of total parameters, number of multiply-accumulate operations (MACs), percentage of total MACs, latency, percentage of total latency, number of floating point operations per second (FLOPS, computed as 2 * MACs / latency).
 Note:
 1. A module can have torch.nn.functional (e.g. to compute logits) along with submodules, thus making the difference between the parent's MACs(or latency) and the sum of its submodules'.
-2. Number of floating point operations is a theoretical estimation, thus FLOPS computed using that could be larger than the maximum system throught.
+2. Number of floating point operations is a theoretical estimation, thus FLOPS computed using that could be larger than the maximum system throughput.
 
 DistributedDataParallel(
   38.89 M, 100.00% Params, 314.61 GMACs, 100.00% MACs, 33.81 ms, 100.00% latency, 18.61 TFLOPS,
@@ -235,11 +235,11 @@ Top 3 modules in params at depth 2 are {'Linear': '58.63 M', 'Conv2d': '2.47 M',
 Top 3 modules in latency at depth 2 are {'Conv2d': '13.96 ms', 'Linear': '6.23 ms', 'ReLU': '730.75 us'}
 
 ------------------------------ Detailed Profile ------------------------------
-Each module profile is listed after its name in the follwing order:
+Each module profile is listed after its name in the following order:
 number of parameters, percentage of total parameters, number of multiply-accumulate operations (MACs), percentage of total MACs, latency, percentage of total latency, number of floating point operations per second (FLOPS, computed as 2 * MACs / latency).
 Note:
 1. A module can have torch.nn.functional (e.g. to compute logits) along with submodules, thus making the difference between the parent's MACs(or latency) and the sum of its submodules'.
-2. Number of floating point operations is a theoretical estimation, thus FLOPS computed using that could be larger than the maximum system throught.
+2. Number of floating point operations is a theoretical estimation, thus FLOPS computed using that could be larger than the maximum system throughput.
 
 AlexNet(
   61.1 M, 100.00% Params, 183.18 GMACs, 100.00% MACs, 22.13 ms, 100.00% latency, 16.56 TFLOPS,
@@ -335,11 +335,11 @@ Top 3 modules in params at depth 7 are {'Linear': '28.35 M', 'LayerNorm': '18.43
 Top 3 modules in latency at depth 7 are {'Linear': '153.7 ms', 'LayerNorm': '4.74 ms', 'Dropout': '597.95 us'}
 
 ------------------------------ Detailed Profile ------------------------------
-Each module profile is listed after its name in the follwing order:
+Each module profile is listed after its name in the following order:
 number of parameters, percentage of total parameters, number of multiply-accumulate operations (MACs), percentage of total MACs, latency, percentage of total latency, number of floating point operations per second (FLOPS, computed as 2 * MACs / latency).
 Note:
 1. A module can have torch.nn.functional (e.g. to compute logits) along with submodules, thus making the difference between the parent's MACs(or latency) and the sum of its submodules'.
-2. Number of floating point operations is a theoretical estimation, thus FLOPS computed using that could be larger than the maximum system throught.
+2. Number of floating point operations is a theoretical estimation, thus FLOPS computed using that could be larger than the maximum system throughput.
 
 BertForSequenceClassification(
   109.48 M, 100.00% Params, 43.5 GMACs, 100.00% MACs, 393.7 ms, 100.00% latency, 220.97 GFLOPS,

docs/_tutorials/pipeline.md

@@ -230,7 +230,7 @@ pipeline. Each worker should load micro-batches of size
 a total of `engine.gradient_accumulation_steps()` times per `train_batch()`.
 
 **Watch out!**
-The pipeline engine *pulls* data from an iteratior instead of iterating over
+The pipeline engine *pulls* data from an iterator instead of iterating over
 it. It's critical that the data stream does not empty in the middle of a
 training batch. Each invocation of `train_batch()` will pull
 a total of `engine.gradient_accumulation_steps()` micro-batches of data from

docs/_tutorials/sparse-attention.md

@@ -154,7 +154,7 @@ This module, is the parent class for all sparsity structures and contains the sh
   * `block`: an integer determining the block size. Current implementation of sparse self-attention is based on blocked sparse matrices. In which this parameter defines size of such square blocks; `Block X Block`.
   * `different_layout_per_head`: a boolean determining if each head should be assigned a different sparsity layout; default is false and this will be satisfied based on availability.
 
-* **Fixed** (FixedSparistyConfig):
+* **Fixed** (FixedSparsityConfig):
 This structure is based on [Generative Modeling with Sparse Transformers](https://arxiv.org/abs/1904.10509) from OpenAI, in which local and global attention is fixed by the given parameters:
   * `num_local_blocks`: an integer determining the number of blocks in local attention window. As it is illustrated in the below figure (adapted from original paper), tokens in a local window, attend to all tokens local to them. In the case of autoregressive model, as in the figure, tokens attend to tokens appearing before them in the local window. And in the case of Masked model such as BERT, attention is bidirectional.
   * `num_global_blocks`: an integer determining how many consecutive blocks in a local window is used as the representative of the window for global attention; illustrated in the figure below as well.

docs/_tutorials/zero.md

@@ -3,7 +3,7 @@ title: "Zero Redundancy Optimizer (ZeRO)"
 ---
 If you have not done so already, we advise that you read the DeepSpeed tutorials on [Getting Started](/getting-started/) and [Megatron-LM GPT-2](/tutorials/megatron/) before stepping through this tutorial.
 
-In this tutorial, we will apply the ZeRO optimizer to the [Megatron-LM GPT-2](https://github.com/NVIDIA/Megatron-LM) model. ZeRO is a powerful set of memory optimization techniques that enable effective FP16 training of large models with trillons of parameters, such as [GPT-2](https://openai.com/blog/better-language-models/) and [Turing-NLG 17B](https://www.microsoft.com/en-us/research/blog/turing-nlg-a-17-billion-parameter-language-model-by-microsoft/). Compared to the alternative model parallelism approaches for training large models, a key appeal of ZeRO is that no model code modifications are required. As this tutorial will demonstrate, *using ZeRO in a DeepSpeed model is quick and easy because all you need is to change a few configurations in the DeepSpeed configuration JSON*. No code changes are needed.
+In this tutorial, we will apply the ZeRO optimizer to the [Megatron-LM GPT-2](https://github.com/NVIDIA/Megatron-LM) model. ZeRO is a powerful set of memory optimization techniques that enable effective FP16 training of large models with trillions of parameters, such as [GPT-2](https://openai.com/blog/better-language-models/) and [Turing-NLG 17B](https://www.microsoft.com/en-us/research/blog/turing-nlg-a-17-billion-parameter-language-model-by-microsoft/). Compared to the alternative model parallelism approaches for training large models, a key appeal of ZeRO is that no model code modifications are required. As this tutorial will demonstrate, *using ZeRO in a DeepSpeed model is quick and easy because all you need is to change a few configurations in the DeepSpeed configuration JSON*. No code changes are needed.
 
 ## ZeRO Overview
 ZeRO leverages the aggregate computation and memory resources of data parallelism to reduce the memory and compute requirements of each device (GPU) used for model training. ZeRO reduces the memory consumption of each GPU by partitioning the various model training states (weights, gradients, and optimizer states) across the available devices (GPUs and CPUs) in the distributed training hardware. Concretely, ZeRO is being implemented as incremental stages of optimizations, where optimizations in earlier stages are available in the later stages. To deep dive into ZeRO, please see our [paper](https://arxiv.org/abs/1910.02054v3).
@@ -226,7 +226,7 @@ class ParallelTransformerLayer(MegatronModule):
 
 #### Allocating Massive Megatron-LM Models
 
-We make two further changes to model initalization in order to support models
+We make two further changes to model initialization in order to support models
 that exceed *local* system memory, but not *total* system memory.
 
 1. Allocate the model in a memory-scalable fashion. The model parameters will