Address Jared's comments in README and loss_scaler.py

README.md

-[Megatron](https://arxiv.org/pdf/1909.08053.pdf) is a large, powerful transformer developed by the Applied Deep Learning Research team at NVIDIA. This repository is for ongoing research on training large transformer language models at scale. We developed efficient, tensor-model-parallel, and multinode training of [GPT-2](https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf) and [BERT](https://arxiv.org/pdf/1810.04805.pdf) using mixed precision.
+[Megatron](https://arxiv.org/pdf/1909.08053.pdf) is a large, powerful transformer developed by the Applied Deep Learning Research team at NVIDIA. This repository is for ongoing research on training large transformer language models at scale. We developed efficient, model-parallel (tensor and pipeline), and multinode training of [GPT-2](https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf) and [BERT](https://arxiv.org/pdf/1810.04805.pdf) using mixed precision.
 
 Using our GPT-2 model we achieve a perplexity of 10.8 on the WikiText-103 dataset (improving SOTA from 15.8) and an accuracy of 66.5% on the LAMBADA datasets. For BERT training, we swapped the position of the layer normalization and the residual connection in the model architecture (similar to GPT-2 architucture), which allowed the models to continue to improve as they were scaled up. Our BERT models with 3.9 billion parameters reaches a loss of 1.16, SQuAD 2.0 F1-score of 91.7, and RACE accuracy of 90.9%.
 
@@ -218,7 +218,7 @@ These scripts use the PyTorch distributed launcher for distributed training. As
 
 The two tiers of parallelism are data and model parallelism. First, we facilitate two distributed data parallel implementations: a simple one of our own that performs gradient all-reduce at the end of back propagation step, and Torch's distributed data parallel wrapper that overlaps gradient reduction with back propagation computation. To switch between these two options use `--DDP-impl local` or `--DDP-impl torch`, respectively. As expected, Torch distributed data parallelism is more efficient at larger model parallel sizes. For example, for the 8.3 billion parameters model running on 512 GPUs, the scaling increases from 60% to 76% when Torch's distributed data parallel is used. However, the overlapping method requires more memory and for some configurations (e.g., 2.5 billion parameters using 2-way model parallel and 1.2 billion parameters with no model parallel) can make the overall training slower as a result. We empirically found that using a smaller model in those cases improves the training time.
 
-Second, we developed a simple and efficient tensor model parallel approach. To use model parallelism, add the `--tensor-model-parallel-size` flag to specify the number of GPUs among which to split the model, along with the arguments passed to the distributed launcher as mentioned above. With `WORLD_SIZE` GPUs and `MP_SIZE` model parallel size, `WORLD_SIZE`/`MP_SIZE` GPUs will be used for data parallelism. The default value for `--tensor-model-parallel-size` is 1, which will not implement model parallelism.
+Second, we developed a simple and efficient two-dimensional model-parallel approach. To use tensor model parallelism (splitting execution of a single transformer module over multiple GPUs), add the `--tensor-model-parallel-size` flag to specify the number of GPUs among which to split the model, along with the arguments passed to the distributed launcher as mentioned above. To use pipeline model parallelism (sharding the transformer modules into stages with an equal number of transformer modules on each stage), use the `--pipeline-model-parallel-size` flag to specify the number of stages to split the model into (e.g., splitting a model with 24 transformer layers across 4 stages would mean each stage gets 6 transformer layers each). With `WORLD_SIZE` GPUs, `TENSOR_MP_SIZE` tensor-model-parallel size, `PIPELINE_MP_SIZE` pipeline-model-parallel-size, `WORLD_SIZE`/(`TENSOR_MP_SIZE` * `PIPELINE_MP_SIZE`) GPUs will be used for data parallelism. The default values for `--tensor-model-parallel-size` and `--pipeline-model-parallel-size` is 1, which will not implement either form of model parallelism.
 
 Other than these minor changes, the distributed training is identical to the training on a single GPU.
 

megatron/fp16/loss_scaler.py

@@ -69,6 +69,10 @@ class LossScaler:
         return grad_in
 
     def backward(self, output_tensor, retain_graph=False, output_tensor_grad=None):
+        # If output_tensor_grad is None, this is the last stage, and
+        # output_tensor is actually the loss and needs to be scaled.
+        # Otherwise, output_tensor does not need to be scaled again since
+        # output_tensor_grad is already scaled.
         if output_tensor_grad is None:
             scaled_output_tensor = output_tensor * self.loss_scale
         else: