LUMI LLM scaling

Scripts and documentation on scaling large language model training on the LUMI supercomputer.

Data

Some example datasets that can be used for the experiments can be found in https://a3s.fi/lumi-llm-scaling/index.html.

For data storage, as the home directory disk size quota on the system is limited to just 20GB, you may want to use one of the other storage options. The simplest option is to set up everything in the project scratch directory /scratch/<project>. For example, if you are a member of project 462000273, you can create a directory for your work in project scratch with

mkdir /scratch/project_462000273/$USER
cd /scratch/project_462000273/$USER

and work there.

Quickstart

To get started training with one of the transformer implementations, see instructions in the relevant subdirectory:

• Megatron-DeepSpeed
• BigCode Megatron-LM
• GPT-NeoX
• (others TODO)

Results

Impact of fused kernels

The figure below shows the impact of switching off one or more of the Megatron-DeepSpeed fused kernels on throughput. In brief summary, only the fused masked softmax kernel has a clear impact on throughput in this experiment.

Impact of --num-workers

A non-zero number of dataloader workers provided an approximately 2% increase in throughput in an experiment using the 33B setup, 32 nodes (256 GCDs), data on scratch (flash had issues at the time)

num-workers	samples/sec	TFLOPs
0           22.1        48.4
1           22.7        49.7
2           22.7        49.6
4           22.8        49.8