Conversation
Well not sure if we can expect much from the 1.3b one. Also their example stuff seems to also invoke a refiner. And they are using unipc flow sampling. edit: the refiner also seems to be 30b https://huggingface.co/robbyant/lingbot-video-moe-30b-a3b/tree/main/refiner |
|
Did you actually get expert-choice MoE working or is this model using token-choice routing? |
The current implementation uses token choice, and so does the official implementation. |
The generation quality of the 30B MoE model isn't very good either. |
Summary
Related Issue / Discussion
N/A
Additional Information
I'm not sure whether there's an issue with my implementation or with the model itself, but the quality is clearly not good.
Checklist