What will I miss out on if I use Accelerate’s Deepspeed integration instead of Deepspeed directly? For example, How can I use MoE in deepspeed over here? Similarly, is every native deepspeed function ported into Accelerate?
What will I miss out on if I use Accelerate’s Deepspeed integration instead of Deepspeed directly?
We try to port all features of deepspeed into Accelerate’s Deepspeed integration. Feel free to submit an issue on accelerate if you don’t see a feature that deepspeed supports.