Reproducing Figures in SC21 Paper

This directory contains some of the scripts that were used to produce the results in the Megatron paper that is to appear at SuperComputing 2021. These scripts use Slurm with the pyxis plugin, but can be modified for other schedulers as well.

Git commit

To replicate these results use Megatron-LM commit: 6985e58938d40ad91ac07b0fddcfad8132e1447e

Setup

All the cluster-dependent variables are in CONFIG.sh. Please update the unspecified values (in angle brackets <...>) before launching any scripts.

Scripts

Below is a list of scripts that can be used to reproduce various figures in our paper:

• run_table_1.sh: Table 1 showing weak-scaling throughput for GPT models ranging from 1 billion to 1 trillion parameters.
• run_figure_11.sh: Figure 11 showing the weak-scaling performance of pipeline parallelism.
• run_figure_12.sh: Figure 12 showing the effect of the interleaved schedule on a 175B GPT model.
• run_figure_13.sh: Figure 13 showing the effect of different degrees of pipeline and tensor model parallelism on a model with 162.2 billion parameters.
• run_figure_14.sh: Figure 14 showing the effect of different degrees of data and pipeline model parallelism on a model with 5.9 billion parameters.
• run_figure_15.sh: Figure 15 showing the effect of different degrees of data and tensor model parallelism on a model with 5.9 billion parameters.
• run_figure_16.sh: Figure 16 showing the effect of microbatch size.
• run_figure_17.sh: Figure 17 showing the effect of activation recomputation.
• run_figure_18.sh: Figure 18 showing the effect of the scatter-gather communication optimization.