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%** |
|