violetch24
commited on
Commit
•
f631890
1
Parent(s):
b0ca7da
Update README.md
Browse filescreate model card
README.md
CHANGED
@@ -1,3 +1,61 @@
|
|
1 |
---
|
|
|
|
|
2 |
license: mit
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
language:
|
3 |
+
- ja
|
4 |
license: mit
|
5 |
+
tags:
|
6 |
+
- ja
|
7 |
+
- japanese
|
8 |
+
- gpt_neox
|
9 |
+
- gpt
|
10 |
+
- text-generation
|
11 |
+
- lm
|
12 |
+
- nlp
|
13 |
+
- int8
|
14 |
+
- neural-compressor
|
15 |
+
- Intel® Neural Compressor
|
16 |
+
- PostTrainingStatic
|
17 |
+
datasets:
|
18 |
+
- oscar
|
19 |
+
model-index:
|
20 |
+
- name: gpt-neox-japanese-2.7b-int8
|
21 |
+
results:
|
22 |
+
- task:
|
23 |
+
name: Text Generation
|
24 |
+
type: text-generation
|
25 |
+
dataset:
|
26 |
+
name: oscar
|
27 |
+
type: oscar
|
28 |
+
args: unshuffled_original_ast
|
29 |
+
metrics:
|
30 |
+
- name: Acurracy
|
31 |
+
type: loss
|
32 |
+
value: 4.9920
|
33 |
---
|
34 |
+
# INT8 gpt-neox-japanese-2.7b-int8
|
35 |
+
|
36 |
+
## Post-training static quantization
|
37 |
+
|
38 |
+
### PyTorch
|
39 |
+
|
40 |
+
This is an INT8 PyTorch model quantized with [Intel® Neural Compressor](https://github.com/intel/neural-compressor).
|
41 |
+
|
42 |
+
The original fp32 model comes from the fine-tuned model [abeja/gpt-neox-japanese-2.7b](https://huggingface.co/abeja/gpt-neox-japanese-2.7b).
|
43 |
+
|
44 |
+
The calibration dataloader is the train dataloader. The default calibration sampling size 100 isn't divisible exactly by batch size 8, so the real sampling size is 104.
|
45 |
+
|
46 |
+
#### Test result
|
47 |
+
|
48 |
+
| |INT8|FP32|
|
49 |
+
|---|:---:|:---:|
|
50 |
+
| **Accuracy (eval-loss)** |4.9920|3.5219|
|
51 |
+
| **Model size (MB)** |2570|5360|
|
52 |
+
|
53 |
+
#### Load with Intel® Neural Compressor:
|
54 |
+
|
55 |
+
```python
|
56 |
+
from optimum.intel.neural_compressor.quantization import IncQuantizedModelForCausalLM
|
57 |
+
|
58 |
+
int8_model = IncQuantizedModelForCausalLM.from_pretrained(
|
59 |
+
"Intel/gpt-neox-japanese-2.7b-int8",
|
60 |
+
)
|
61 |
+
```
|