Commit
•
5f4af14
1
Parent(s):
8a9d3f4
Update README.md (#1)
Browse files- Update README.md (28fad366da38c31a490977e9f86ae43df0ced4a0)
Co-authored-by: Abhinav Agarwalla <abhinavnmagic@users.noreply.huggingface.co>
README.md
CHANGED
|
@@ -4,11 +4,13 @@ tags:
|
|
| 4 |
- vllm
|
| 5 |
---
|
| 6 |
|
| 7 |
-
# Qwen2-72B-Instruct-FP8
|
| 8 |
|
| 9 |
-
|
|
|
|
| 10 |
|
| 11 |
-
|
|
|
|
| 12 |
|
| 13 |
```python
|
| 14 |
from datasets import load_dataset
|
|
@@ -33,4 +35,19 @@ model = AutoFP8ForCausalLM.from_pretrained(
|
|
| 33 |
)
|
| 34 |
model.quantize(examples)
|
| 35 |
model.save_quantized(quantized_model_dir)
|
| 36 |
-
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
- vllm
|
| 5 |
---
|
| 6 |
|
| 7 |
+
# Qwen2-72B-Instruct-FP8
|
| 8 |
|
| 9 |
+
## Model Overview
|
| 10 |
+
Qwen2-72B-Instruct quantized to FP8 weights and activations using per-tensor quantization, ready for inference with vLLM >= 0.5.0.
|
| 11 |
|
| 12 |
+
## Usage and Creation
|
| 13 |
+
Produced using [AutoFP8 with calibration samples from ultrachat](https://github.com/neuralmagic/AutoFP8/blob/147fa4d9e1a90ef8a93f96fc7d9c33056ddc017a/example_dataset.py).
|
| 14 |
|
| 15 |
```python
|
| 16 |
from datasets import load_dataset
|
|
|
|
| 35 |
)
|
| 36 |
model.quantize(examples)
|
| 37 |
model.save_quantized(quantized_model_dir)
|
| 38 |
+
```
|
| 39 |
+
|
| 40 |
+
## Evaluation
|
| 41 |
+
|
| 42 |
+
### Open LLM Leaderboard evaluation scores
|
| 43 |
+
| | Qwen2-72B-Instruct | Qwen2-72B-Instruct-FP8<br>(this model) |
|
| 44 |
+
| :------------------: | :----------------------: | :------------------------------------------------: |
|
| 45 |
+
| arc-c<br>25-shot | 71.58 | 72.09 |
|
| 46 |
+
| hellaswag<br>10-shot | 86.94 | 86.83 |
|
| 47 |
+
| mmlu<br>5-shot | xx.xx | 84.06 |
|
| 48 |
+
| truthfulqa<br>0-shot | 66.94 | 66.95 |
|
| 49 |
+
| winogrande<br>5-shot | 82.79 | 83.18 |
|
| 50 |
+
| gsm8k<br>5-shot | xx.xx | 88.93 |
|
| 51 |
+
| **Average<br>Accuracy** | **xx.xx** | **80.34** |
|
| 52 |
+
| **Recovery** | **100%** | **xx.xx%** |
|
| 53 |
+
|