Deepseek-v3-Base Group 8 Average Weights

Since Deepseek v3 dense layers ( first 3 ) happens to be 18432 which equals to 9 x 2048. Since there's 256 experts with 2048 dimensions as the intermediate dimension, we can first average all 32 experts as 1 expert and concatenate all 8 groups into a 16384 layers. Adding the share_experts in to the new MLP layers and we can get 18432 MLP layers.

Model Details

Model Description

Unfortunately, this model doesn't work out of the box ( after dequantize and merging ) all it generates is giberish tokens. So either my code sucks or merging all that experts down breaks the model too much that every brokes.

I'm trying to recover the MLP layers by pretraining, but I'm bit low on compute and doesn't have much to spare. Also if you have small corpus which I can use feel free to comments and suggest what I should do next

QLoRA pretrained test run can be found here : theblackcat102/whale-v3-base-concept-test-lora-380