monsoon-nlp
commited on
Commit
•
e39d68d
1
Parent(s):
882fa80
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,62 @@
|
|
1 |
---
|
2 |
license: apache-2.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
+
language:
|
4 |
+
- ar
|
5 |
+
- hi
|
6 |
+
- id
|
7 |
+
pipeline_tag: text-generation
|
8 |
+
tags:
|
9 |
+
- multilingual
|
10 |
+
widget:
|
11 |
+
- text: 'في مدرستي السابقة'
|
12 |
+
example_title: Arabic prompt
|
13 |
+
- text: 'आप समुद्री लुटेरों के बारे में क्या जानते हैं?'
|
14 |
+
example_title: Hindi prompt
|
15 |
+
- text: 'Kucing saya suka'
|
16 |
+
example_title: Indonesian prompt
|
17 |
---
|
18 |
+
|
19 |
+
# mGPT-quantized
|
20 |
+
|
21 |
+
The concept: 8-bit quantized version of [mGPT-13B](https://huggingface.co/ai-forever/mGPT-13B), an LLM released by AI-Forever / Sberbank AI in 2022-2023.
|
22 |
+
|
23 |
+
On the GPT scale, it is a similar # of parameters to GPT-3, but trained on 60+ languages.
|
24 |
+
|
25 |
+
My goal is to evaluate this on Hindi and Indonesian tasks, where there are fewer autoregressive language models in this size range.
|
26 |
+
|
27 |
+
For English: use a GPT model or LLaMa2-7B
|
28 |
+
|
29 |
+
For Arabic: in August 2023 I would recommend the bilingual [JAIS model](https://huggingface.co/inception-mbzuai/jais-13b), which is also 13B parameters can be quantized.
|
30 |
+
|
31 |
+
In August 2023 AI-Forever added 1.3B-param models for 20+ languages. If your language is Mongolian, for example, it might be better to use mGPT-1.3B-mongol and not this one.
|
32 |
+
|
33 |
+
They also have a 1.3B param model for all languages, which I further quantized here: https://huggingface.co/monsoon-nlp/mGPT-quantized
|
34 |
+
|
35 |
+
## How was the model created?
|
36 |
+
|
37 |
+
Quantization of mGPT-13B was done using `bitsandbytes` library, CoLab Pro with an A100 GPU, and a lot of space on Google Drive.
|
38 |
+
|
39 |
+
```python
|
40 |
+
from transformers import BitsAndBytesConfig, GPT2LMHeadModel
|
41 |
+
|
42 |
+
quantization_config = BitsAndBytesConfig(
|
43 |
+
load_in_8bit=True,
|
44 |
+
bnb_8bit_compute_dtype=torch.bfloat16,
|
45 |
+
bnb_8bit_use_double_quant=True,
|
46 |
+
bnb_8bit_quant_type="nf4",
|
47 |
+
)
|
48 |
+
|
49 |
+
qmodel = GPT2LMHeadModel.from_pretrained(
|
50 |
+
"ai-forever/mGPT-13B",
|
51 |
+
load_in_8bit=True,
|
52 |
+
torch_dtype=torch.bfloat16,
|
53 |
+
quantization_config=quantization_config,
|
54 |
+
device_map="auto"
|
55 |
+
)
|
56 |
+
|
57 |
+
qmodel.save_pretrained("model_name")
|
58 |
+
```
|
59 |
+
|
60 |
+
## Future steps
|
61 |
+
|
62 |
+
- mGPT could be further quantized (4-bit), but `model.save_pretrained()` currently throws a `NotImplementedError` error.
|