From a1815330a4bd2b51542399cc3dd1336ce98a445c Mon Sep 17 00:00:00 2001 From: Tongjilibo Date: Thu, 18 Apr 2024 11:52:33 -0400 Subject: [PATCH] v0.5.0 --- README.md | 3 +-- docs/History.md | 2 +- docs/Update.md | 2 ++ examples/llm/task_chatglm_lora.py | 2 +- examples/sequence_labeling/uie/README.md | 2 +- setup.py | 4 ++-- 6 files changed, 8 insertions(+), 7 deletions(-) diff --git a/README.md b/README.md index aded0f07..632490e3 100644 --- a/README.md +++ b/README.md @@ -93,10 +93,9 @@ pip install git+https://github.com/Tongjilibo/bert4torch ### 4.1 版本历史 |更新日期| bert4torch | torch4keras | 版本说明 | |------| ---------------- | ----------------- |----------- | +|20240418| 0.5.0 | 0.2.2 | 修复chatglm3的bug, 修复save_pretrained时多文件的bug，增加CausalLMLoss, 修改deepspeed的传参逻辑，修改Text2Vec的bug, 完善openai client, 增加get_weight_decay_optim_groups| |20240317| 0.4.9.post2 | 0.2.1.post2 |增加get_weight_decay_optim_groups函数, attention中允许is_causal，修改repetition_penalty的bug，把baichuan从llama中剥离，修复config_path的bug，允许num_key_value_heads参数，[torch4keras-v0.2.1.post2](https://github.com/Tongjilibo/torch4keras/releases/tag/v0.2.1.post2)更新特性| |20240221| 0.4.8 | 0.2.0|fastapi发布服务允许闲时offload到cpu, `build_transformer_model`允许从hf下载, 添加`FillMask`的pipeline, 添加`SequenceClassificationTrainer`| -|20240204| 0.4.7 | 0.1.9|修改`save_pretrained`用于保存文件夹, 增加GenerateSpeed用于统计token生成速度，修复t5在use_states=True时候的错误, 修改层次编码的bug, 增加deepseek_moe模型，修复generation并发错误，优化大模型耗时| -|20240116| 0.4.6 | 0.1.8|bug修复，增加`save_pretrained`用于保存`transformer`格式的权重, 增加部分`embedding`模型| [更多版本](https://github.com/Tongjilibo/bert4torch/blob/master/docs/Update.md) diff --git a/docs/History.md b/docs/History.md index c8cbe9ff..c1d47a62 100644 --- a/docs/History.md +++ b/docs/History.md @@ -1,6 +1,6 @@ ## 更新历史 -- **20240403**：修改Text2Vec的bug +- **20240418**：修改Text2Vec的bug, 完善openai client, 增加get_weight_decay_optim_groups - **20240331**: 修复chatglm3的bug, 修复save_pretrained时多文件的bug，增加CausalLMLoss, 修改deepspeed的传参逻辑 - **20240317**: 修复config_path的bug，允许num_key_value_heads参数 - **20240316**: 增加get_weight_decay_optim_groups函数, attention中允许is_causal，修改repetition_penalty的bug，把baichuan从llama中剥离，[torch4keras-v0.2.1](https://github.com/Tongjilibo/torch4keras/releases/tag/v0.2.1)更新特性 diff --git a/docs/Update.md b/docs/Update.md index b7f3cfa8..b60a3e4f 100644 --- a/docs/Update.md +++ b/docs/Update.md @@ -2,6 +2,8 @@ |更新日期| bert4torch版本 | torch4keras版本 | 版本说明 | |------| ---------------- | ----------------- |----------- | +|20240418| 0.5.0 | 0.2.2 | 修复chatglm3的bug, 修复save_pretrained时多文件的bug，增加CausalLMLoss, 修改deepspeed的传参逻辑，修改Text2Vec的bug, 完善openai client, 增加get_weight_decay_optim_groups| +|20240317| 0.4.9.post2 | 0.2.1.post2 |增加get_weight_decay_optim_groups函数, attention中允许is_causal，修改repetition_penalty的bug，把baichuan从llama中剥离，修复config_path的bug，允许num_key_value_heads参数，[torch4keras-v0.2.1.post2](https://github.com/Tongjilibo/torch4keras/releases/tag/v0.2.1.post2)更新特性| |20240221| 0.4.8 | 0.2.0|fastapi发布服务允许闲时offload到cpu, `build_transformer_model`允许从hf下载, 添加`FillMask`的pipeline, 添加`SequenceClassificationTrainer`| |20240204| 0.4.7 | 0.1.9|修改`save_pretrained`用于保存文件夹, 增加GenerateSpeed用于统计token生成速度，修复t5在use_states=True时候的错误, 修改层次编码的bug, 增加deepseek_moe模型，修复generation并发错误，优化大模型耗时| |20240116| 0.4.6 | 0.1.8|bug修复，增加`save_pretrained`用于保存`transformer`格式的权重, 增加部分`embedding`模型| diff --git a/examples/llm/task_chatglm_lora.py b/examples/llm/task_chatglm_lora.py index c16ee1d3..e503d244 100644 --- a/examples/llm/task_chatglm_lora.py +++ b/examples/llm/task_chatglm_lora.py @@ -41,7 +41,7 @@ max_target_length = 64 max_seq_length = max_source_length + max_target_length lr = 5e-4 -batch_size = 4 # 根据显存大小调整 +batch_size = 16 # 根据显存大小调整 eval_batch_size = 4 grad_accumulation_steps = 1 # 根据显存大小调整 epochs = 1 diff --git a/examples/sequence_labeling/uie/README.md b/examples/sequence_labeling/uie/README.md index 51fa75e5..c2b8f55f 100644 --- a/examples/sequence_labeling/uie/README.md +++ b/examples/sequence_labeling/uie/README.md @@ -1,5 +1,5 @@ # 说明 -- convert.py: 下载uie模型并且转成pytorch格式 +- convert.py: 下载uie模型并且转成pytorch格式，作者已转换的[权重](https://huggingface.co/Tongjilibo/uie-base) - finetune.sh：数据处理和训练全流程 - finetune_step1_dataprocess.py：数据预处理1 - finetune_step2_doccano.py：数据预处理2 diff --git a/setup.py b/setup.py index 03d41fa4..8ce40af3 100644 --- a/setup.py +++ b/setup.py @@ -7,13 +7,13 @@ setup( name='bert4torch', - version='v0.4.9.post2', + version='v0.5.0', description='an elegant bert4torch', long_description=long_description, long_description_content_type="text/markdown", license='MIT Licence', url='https://github.com/Tongjilibo/bert4torch', author='Tongjilibo', - install_requires=['numpy', 'tqdm', 'torch>1.6', 'torch4keras==0.2.1.post2', 'six'], + install_requires=['numpy', 'tqdm', 'torch>1.6', 'torch4keras==0.2.2', 'six'], packages=find_packages() ) \ No newline at end of file