File size: 1,600 Bytes
185a29e
 
 
8b22b46
185a29e
5daf216
8b22b46
09cfe83
8b22b46
 
 
 
d288bbc
8b22b46
 
 
 
 
 
28b8603
 
54e6f9f
28b8603
 
 
54e6f9f
 
8b22b46
 
 
 
 
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
---
tags:
- fp8
- vllm
---

# Mixtral-8x7B-Instruct-v0.1-FP8

## Model Overview
Mixtral-8x7B-Instruct-v0.1 quantized to FP8 weights and activations, ready for inference with vLLM >= 0.5.0.

## Usage and Creation
Produced using [AutoFP8 with calibration samples from ultrachat](https://github.com/neuralmagic/AutoFP8/blob/147fa4d9e1a90ef8a93f96fc7d9c33056ddc017a/examples/example_mixtral.py) with `block_sparse_moe.gate` layers kept at original precision.

## Evaluation

### Open LLM Leaderboard evaluation scores
|                      | Mixtral-8x7B-Instruct-v0.1 | Mixtral-8x7B-Instruct-v0.1-FP8<br>(this model) |
| :------------------: | :----------------------: | :------------------------------------------------: |
| arc-c<br>25-shot     | 71.50                    | 71.08                                              |
| hellaswag<br>10-shot | 87.53                    | 87.38                                              |
| mmlu<br>5-shot       | 70.33                    | 70.00                                              |
| truthfulqa<br>0-shot | 64.79                    | 64.20                                              |
| winogrande<br>5-shot | 82.40                    | 82.40                                              |
| gsm8k<br>5-shot      | 64.36                    | 64.06                                              |
| **Average<br>Accuracy**  | **73.48**                    |              **73.19**                                     |
| **Recovery**             | **100%**                     |              **99.61%**                                     |