Update README.md
Browse files
README.md
CHANGED
|
@@ -1,19 +1,33 @@
|
|
| 1 |
---
|
| 2 |
tags:
|
| 3 |
-
- awq
|
| 4 |
-
- llm
|
| 5 |
-
- quantization
|
| 6 |
---
|
| 7 |
|
| 8 |
# yujiepan/awq-model-zoo
|
| 9 |
|
| 10 |
Here are some pre-computed awq information (scales & clips) used in [llm-awq](https://github.com/mit-han-lab/llm-awq).
|
| 11 |
|
| 12 |
-
|
| 13 |
## Scripts
|
| 14 |
|
| 15 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
|
| 18 |
## Related links
|
| 19 |
-
|
|
|
|
|
|
| 1 |
---
|
| 2 |
tags:
|
| 3 |
+
- awq
|
| 4 |
+
- llm
|
| 5 |
+
- quantization
|
| 6 |
---
|
| 7 |
|
| 8 |
# yujiepan/awq-model-zoo
|
| 9 |
|
| 10 |
Here are some pre-computed awq information (scales & clips) used in [llm-awq](https://github.com/mit-han-lab/llm-awq).
|
| 11 |
|
|
|
|
| 12 |
## Scripts
|
| 13 |
|
| 14 |
+
- Install the forked `llm-awq` at [https://github.com/yujiepan-work/llm-awq/tree/a41a08e79d8eb3d6335485b3625410af22a74426](https://github.com/yujiepan-work/llm-awq/tree/a41a08e79d8eb3d6335485b3625410af22a74426). Note: works with transformers==4.35.2
|
| 15 |
+
|
| 16 |
+
- Generating awq-info.pt:
|
| 17 |
+
|
| 18 |
+
```bash
|
| 19 |
+
python do_awq.py --model_id mistralai/Mistral-7B-v0.1 --w_bit 8 --q_group_size 128 --dump_awq ./awq-info.pt
|
| 20 |
+
```
|
| 21 |
|
| 22 |
+
- Load a quantized model: You can use the offical repo to get a fake/real quantized model. Alternatively, you can load a fake-quantized model:
|
| 23 |
+
|
| 24 |
+
```python
|
| 25 |
+
from do_awq import FakeAWQModel
|
| 26 |
+
FakeAWQModel.from_pretrained('mistralai/Mistral-7B-v0.1', awq_meta_path='./awq-info.pt', output_folder='./tmp/')
|
| 27 |
+
```
|
| 28 |
+
|
| 29 |
+
Note: the code is not in good shape.
|
| 30 |
|
| 31 |
## Related links
|
| 32 |
+
|
| 33 |
+
- <https://huggingface.co/datasets/mit-han-lab/awq-model-zoo>
|