|
--- |
|
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%** | |
|
|
|
|
|
|
|
|
|
|
|
|
