mirror of
https://gitee.com/mindspore/mindformers.git
synced 2025-12-06 11:29:59 +08:00
更新qwen2.5-72B-32K性能yaml,启用dw掩盖提升性能
This commit is contained in:
@@ -150,11 +150,12 @@ context:
|
||||
device_id: 0
|
||||
ascend_config:
|
||||
precision_mode: "must_keep_origin_dtype"
|
||||
parallel_speed_up_json_path: "./research/qwen2_5/parallel_speed_up_72B_32K.json" # Replace with a real path when needed
|
||||
|
||||
# model config
|
||||
model:
|
||||
model_config:
|
||||
use_attn_mask_compression: True
|
||||
use_attn_mask_compression: False
|
||||
type: LlamaConfig
|
||||
batch_size: 1
|
||||
seq_length: 32768
|
||||
|
||||
3
research/qwen2_5/parallel_speed_up_72B_32K.json
Normal file
3
research/qwen2_5/parallel_speed_up_72B_32K.json
Normal file
@@ -0,0 +1,3 @@
|
||||
{
|
||||
"matmul_grad_comm_overlap": true
|
||||
}
|
||||
@@ -227,7 +227,23 @@ train_dataset: &train_dataset
|
||||
dynamic_batch: True
|
||||
```
|
||||
|
||||
4. 启动微调:
|
||||
4. 配置并行加速
|
||||
|
||||
若模型的yaml中有类似如下的配置
|
||||
|
||||
```yaml
|
||||
context:
|
||||
ascend_config:
|
||||
parallel_speed_up_json_path: "/path/to/parallel_speed_up.json" # Replace with a real path when needed
|
||||
```
|
||||
|
||||
这是在使用`parallel_speed_up`文件(须是`json`格式)去配置一些并行加速特性,以获得一些性能上的提升。实际使用时,请把`json`文件的路径修改为实际值。
|
||||
|
||||
> `parallel_speed_up`文件中各配置项的含义详见[parallel_speed_up说明](https://www.mindspore.cn/docs/zh-CN/master/api_python/mindspore/mindspore.set_context.html#:~:text=parallel_speed_up_json_path)。
|
||||
|
||||
例如,`finetune_qwen2.5_72B_32K.yaml`中使用了`parallel_speed_up_72B_32K.json`,其中配置了`"matmul_grad_comm_overlap": true`。
|
||||
|
||||
5. 启动微调:
|
||||
|
||||
```shell
|
||||
cd research/qwen2_5
|
||||
|
||||
Reference in New Issue
Block a user