metadata

language:
  - en
license: apache-2.0
library_name: transformers
tags:
  - moe
  - moah
  - mod
  - mh-moe
datasets:
  - Locutusque/UltraTextbooks

Model Card for Model ID

Model Details

MoM: Mixture of Mixture

This Model is a first test to combine Jamba architecture with bf16 bits linear layers, mixture of attention head and multi head mixture of depth.

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 bf16 linear layers
License: Apache licence 2.0

This model has a generation problem because of a softmax application in the mod process

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 metadata to see the hyperparameters or the train.py file in the repo