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

# Model Card for Model ID

## Model Details

### Model Description

<!-- Provide a longer summary of what this model is. -->

MoM: Mixture of Mixture

This Model is a first test to combine [Jamba](https://huggingface.co/ai21labs/Jamba-v0.1) 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

### Model Sources [optional]


- **Repository:** https://github.com/ostix360/optimized-LLM


## How to Get Started with the Model


If you want to test  this model please look at this repo at this [commit](https://github.com/ostix360/optimized-LLM/tree/796cfe43cf16461b92102cf0f41e8960cd91340b)


## Training Details

  - **wandb**: [training detail](https://wandb.ai/ostix360/Mixture%20of%20mixture%20(mod,%20moah%20moe)/runs/0ayclh2i)

### Training Data

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

### Training Procedure

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

#### Preprocessing [optional]


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


#### Training Hyperparameters

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


## Technical Specifications [optional]

### Compute Infrastructure

#### Hardware

- one 4070 ti GPU 

#### Software

- pytorch, transformers etc