File size: 1,873 Bytes
b40404a
7921f0f
 
 
b40404a
b6466f5
 
 
 
 
 
b40404a
 
 
 
 
 
 
 
 
 
b6466f5
b40404a
b6466f5
b40404a
b6466f5
b40404a
b6466f5
b40404a
b6466f5
b40404a
 
b6466f5
 
b40404a
b6466f5
b40404a
 
b6466f5
b40404a
 
 
 
 
b6466f5
 
b40404a
 
 
b6466f5
b40404a
b6466f5
b40404a
b6466f5
b40404a
 
 
b6466f5
b40404a
 
 
 
b6466f5
b40404a
 
b6466f5
b40404a
b6466f5
b40404a
 
 
 
 
 
 
 
b6466f5
b40404a
 
 
b6466f5
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
---
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.

Mamba and attention layers are in bf16 precision and the rest is in 1.58bits precision

107M over a total of 1025M parameters are in bf16 precision ~ 10% 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 for **MLP** 
- **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/d266bc404346b71ea237c0744be0f8928f6b3217)


## Training Details

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

### Training Data

We use the first 100k data 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