File tree Expand file tree Collapse file tree 2 files changed +15
-3
lines changed Expand file tree Collapse file tree 2 files changed +15
-3
lines changed Original file line number Diff line number Diff line change @@ -193,7 +193,13 @@ accelerate launch --config_file accelerate_fsdp_config.yaml pefts/mft_accelerate
193
193
or
194
194
FSDP config in command line arguments
195
195
``` bash
196
- sh ds_single_launch.sh
196
+ sh fsdp_single_launch.sh
197
+ ```
198
+
199
+ #### MultiNode Launch
200
+ Refer to the deepspeed multi-node launch script below.
201
+ ``` bash
202
+ sh ds_multinode_launch.sh
197
203
```
198
204
199
205
#### Traing Arguments
Original file line number Diff line number Diff line change @@ -146,7 +146,7 @@ QLoRA通过4-bit的nf4量化,且加入更多adapter,在大幅减少显存消
146
146
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。
147
147
148
148
执行如下命令即可进行 Lora/QLora/全量 微调:
149
- #### Launch via Deepspeed
149
+ #### Deepspeed 单机启动
150
150
DeepSpeed配置在accelerate_ds_config.yaml中。
151
151
``` bash
152
152
accelerate launch --config_file accelerate_ds_config.yaml pefts/mft_accelerate.py --train_config configs/xxx_train_config.json --distributed_type " DeepSpeed"
@@ -163,7 +163,7 @@ DeepSpeed Zero3 配置在脚本中通过命令行输入
163
163
sh ds_zero3_single_launch.sh
164
164
```
165
165
166
- #### Launch via FSDP
166
+ #### FSDP 单机启动
167
167
FSDP配置在accelerate_fsdp_config.yaml中。
168
168
``` bash
169
169
accelerate launch --config_file accelerate_fsdp_config.yaml pefts/mft_accelerate.py --train_config configs/xxx_train_config.json --distributed_type " FSDP"
@@ -175,6 +175,12 @@ FSDP配置在脚本中通过命令行输入。
175
175
sh fsdp_single_launch.sh
176
176
```
177
177
178
+ #### 多机启动
179
+ 多机启动请参考如下deepspeed多机启动脚本
180
+ ``` bash
181
+ sh ds_multinode_launch.sh
182
+ ```
183
+
178
184
#### 训练参数
179
185
_ ** 训练需要的参数配置在``` configs/*_train_config ``` 中,主要参数说明如下:** _
180
186
You can’t perform that action at this time.
0 commit comments