Skip to content

Commit

Permalink
Update README.md
Browse files Browse the repository at this point in the history
  • Loading branch information
pass-lin authored Mar 17, 2024
1 parent ac3843d commit 1b925d3
Showing 1 changed file with 6 additions and 1 deletion.
7 changes: 6 additions & 1 deletion README.md
Original file line number Diff line number Diff line change
Expand Up @@ -59,7 +59,7 @@ https://github.com/keras-team/keras/blob/master/requirements-torch-cuda.txt
- <strong>GAU-ALPHA</strong>: https://github.com/ZhuiyiTechnology/GAU-alpha

### bert4keras3的新增加的模型权重,不再使用ckpt存储。通过model.load_weights方法读取权重,建议使用keras3加载
### bert4keras3的新增加的模型权重,不再使用ckpt存储。通过build_transformer_model( keras_weights_path='xx.weights.h5')方法读取权重,只能使用keras3加载
- <strong>自行转换的ChatYuan Keras3版本</strong>:链接:https://pan.baidu.com/s/1huwRA2sOIYL8e7OFg0QTcg?pwd=6u59 提取码:6u59

<strong>注意事项</strong>
Expand All @@ -75,3 +75,8 @@ https://github.com/keras-team/keras/blob/master/requirements-torch-cuda.txt
[roformer/roformer-v2](https://github.com/pass-lin/bert4keras3/blob/main/examples/test_simroformer.py)的cache支持,用法和t5一样,example里只是测试一下与greedy是否一致


## 3.17号更新,发布1.2版本
增加了对weights.h5的读取支持
增加了lora支持,可以通过设置os.environ["ENABLE_LORA"]='1' 启动lora训练,注意的是除了lora之外的参数全部会被冻结
增加了flash-attention支持,可以通过设置os.environ["FLASH_ATTN"]='1'使用flash-attention
但是需要注意的是,tensorflow不支持。而jax在https://github.com/nshepperd/flash_attn_jax/releases下载,torch则是https://github.com/Dao-AILab/flash-attention

0 comments on commit 1b925d3

Please sign in to comment.