Model Card for Model ID

Model Details

MoM: Mixture of Mixture

This Model is a first test to combine Jamba architecture with mixture of attention head and mixture of depth.

Attention layers only are in bf16 precision and the rest is in 1.58bits precision

17M over a total of 1025M parameters are in bf16 precision ~ 1.7% of the parameters are in bf16

The goal is to developpe and test if this kind of architectures have not too much quality loss for a fast inference.

Model type: Mixture of attention head mixture of depth and mixture of expert with 1.58bits linear layer excpeted for attention
License: Apache licence 2.0

If you want to test this model please look at this repo at this commit

We use the first ~0.5B tokens of Locutusque/UltraTextbooks to train this model

We use adam-8 bits with default betas and epsilon values

The data fit the model max length i.e. 512 tokens

Please look at the wandb meta data or the train.py in the repo to see the hyperparameters