|
--- |
|
tags: |
|
- awq |
|
- llm |
|
- quantization |
|
--- |
|
|
|
# yujiepan/awq-model-zoo |
|
|
|
Here are some pre-computed awq information (scales & clips) used in [llm-awq](https://github.com/mit-han-lab/llm-awq). |
|
|
|
## Scripts |
|
|
|
- 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 |
|
|
|
- Generating awq-info.pt: |
|
|
|
```bash |
|
python do_awq.py --model_id mistralai/Mistral-7B-v0.1 --w_bit 8 --q_group_size 128 --dump_awq ./awq-info.pt |
|
``` |
|
|
|
- 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: |
|
|
|
```python |
|
from do_awq import FakeAWQModel |
|
FakeAWQModel.from_pretrained('mistralai/Mistral-7B-v0.1', awq_meta_path='./awq-info.pt', output_folder='./tmp/') |
|
``` |
|
|
|
Note: the code is not in good shape. |
|
|
|
## Related links |
|
|
|
- <https://huggingface.co/datasets/mit-han-lab/awq-model-zoo> |
|
|
