Hi! MegaBlocks isn't a standalone training framework, but it's relatively easy to use from any framework. We use Megatron-LM and have a fork of it with MegaBlocks support. You could also use MegaBlocks from another framework like HuggingFace, for example.
Do you have SFT scripts? And Hyperparameters that you used to fine-tune the instruct version of your model? It would mean the world for the OS community!
Hi! MegaBlocks isn't a standalone training framework, but it's relatively easy to use from any framework. We use Megatron-LM and have a fork of it with MegaBlocks support. You could also use MegaBlocks from another framework like HuggingFace, for example.