Skip to content

Commit 484a346

Browse files
committed
readme
1 parent f9669ce commit 484a346

File tree

4 files changed

+26
-3
lines changed

4 files changed

+26
-3
lines changed

.gitignore

Lines changed: 4 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,2 +1,5 @@
11
.idea/
2-
.DS_Store
2+
.DS_Store
3+
*.log
4+
*/__pycache__/
5+
*.pyc

mftcoder_accelerate/README.md

Lines changed: 9 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -135,6 +135,15 @@ cd mftcoder_accelerate/src
135135

136136
accelerate launch --config_file accelerate_ds_config.yaml pefts/mft_accelerate.py --train_config configs/lora_train_config.json
137137
```
138+
OR
139+
140+
You can launch the training by:
141+
```bash
142+
cd mftcoder_accelerate/src
143+
144+
sh ds_single_launch.sh
145+
```
146+
138147
All arguments allowed in ***_train_config.josn are defined in ```arguments.py```.
139148

140149
Frequently used arguments are provided in ```configs/***_train_config``` and explained as follows. You can modify these parameters according to your needs:

mftcoder_accelerate/README_cn.md

Lines changed: 13 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -116,12 +116,23 @@ cd mftcoder_accelerate/src
116116
QLoRA通过4-bit的nf4量化,且加入更多adapter,在大幅减少显存消耗的同时,尽可能逼近全量参数微调的效果。
117117
QLoRA论文指出,该方法可以在一张V100上对33B的模型进行微调,并且性能逼近全量参数微调。
118118

119-
执行如下命令即可进行Lora/QLora微调:
119+
执行如下命令即可进行Lora/QLora/全量 微调:
120+
121+
deepspeed配置在accelerate_ds_config.yaml中。
120122
```bash
121123
accelerate launch --config_file accelerate_ds_config.yaml pefts/mft_accelerate.py --train_config configs/xxx_train_config.json
122124
```
125+
或者
126+
127+
修改并执行如下sh脚本:
128+
129+
deepspeed配置在脚本中通过命令行输入。
130+
```bash
131+
sh ds_single_launch.sh
132+
```
133+
134+
_**训练需要的参数配置在```configs/*_train_config```中,主要参数说明如下:**_
123135

124-
```configs/*_train_config```中的主要参数说明如下,以下参数可以根据需求进行修改,其他参数建议不做修改:
125136
- load_raw_dataset : 需要保持true,后续会支持其它模式数据,当前仅支持jsonl输入
126137
- data_paths: "[path1,path2,path3]" 输入数据地址,字符串,开头结尾用[],中间用```,```间隔不同path,每个path是一个目录,目录的最后一级名字作为任务名称,下面包含1到多个jsonl数据
127138
- output_dir:训练输出目录,存储checkpoint、lora_adaptor等

0 commit comments

Comments
 (0)