ColossalAI
ColossalAI copied to clipboard
[FEATURE]: Is hybrid parallelism supported in GPT demo?
Describe the feature
I found only DP and ZeRO strategy supports in ColossalAI/applications/Chat/examples, is hybrid parallelism (like PP / Megatron) supported?
If you are looking for examples of training GPT2, there is one solution with our interesting auto-parallelism tool.
Hi @nrailgun DP+ZERO is a simple way and enough for most cases. Examples of hybrid parallelism for large models will be provided soon. Thanks.