Commit
•
b4239f2
1
Parent(s):
4ee4162
Update README.md
Browse files
README.md
CHANGED
|
@@ -22,13 +22,14 @@ Please notice the kernel is not optimzed for 1-bit matrix yet.
|
|
| 22 |
|
| 23 |
The model is trained on a 3090(24GB) for 16 hours.
|
| 24 |
|
|
|
|
| 25 |
### For training code, check https://github.com/Mrw33554432/Bitlinear4HF.
|
| 26 |
|
| 27 |
The training code should be compatible with most of the LLMs in huggingface.
|
| 28 |
|
| 29 |
Using pretrained model weight (normal models) for training will not work due to gradient explosion.
|
| 30 |
|
| 31 |
-
## Sample inference code
|
| 32 |
|
| 33 |
|
| 34 |
```python
|
|
|
|
| 22 |
|
| 23 |
The model is trained on a 3090(24GB) for 16 hours.
|
| 24 |
|
| 25 |
+
### For faster(3x) inference, check https://github.com/Mrw33554432/Bitlinear4HF and install custom kernel
|
| 26 |
### For training code, check https://github.com/Mrw33554432/Bitlinear4HF.
|
| 27 |
|
| 28 |
The training code should be compatible with most of the LLMs in huggingface.
|
| 29 |
|
| 30 |
Using pretrained model weight (normal models) for training will not work due to gradient explosion.
|
| 31 |
|
| 32 |
+
## Sample inference code (slow)
|
| 33 |
|
| 34 |
|
| 35 |
```python
|