add cpm

Author: Lin

.jenkins/test/config/dependent_packages.yaml

@@ -1,2 +1,2 @@
 mindspore:
-  '/mindspore/mindspore/version/202306/20230629/master_20230629013105_f5db72fb1154d12218925ececd304351774f53cc/'
+  '/mindspore/mindspore/version/202307/20230731/r2.1_20230731201928_db56d4dffe0f593f9e983142169b74a4555058db/'

README.md

@@ -146,6 +146,8 @@ configs统一在run_xxx.yaml中，排序按照修改频率的顺序和一般的
         - prefix: 权重文件前缀
         - save_checkpoint_steps: 每多少个step保存一次checkpoint
         - integrated_save: 是否聚合保存。True时表示聚合所有卡权重，这时每张卡权重均一致；False时表示每张卡各自保存自己的权重；当半自动并行模式训练大模型时，通常需要设置为False，以保证权重保存时不会因为内存问题而失败
+        - save_network_params（新增）: 是否额外保存瘦身后的权重。默认为True。
+        - save_trainable_params（新增）: 是否额外保存可训练的参数权重，即微调部分参数的权重。默认为False。
         - async_save: 是否异步执行保存checkpoint文件
     - type: ObsMonitor: obs数据上传
         - step_upload_frequence: 每隔多少个step上传一次，默认为-1，表示禁用step间隔上传；配置为大于0的数时，每隔配置数step后执行一次回传

configs/README.md

@@ -2,6 +2,7 @@ seed: 0
 run_mode: 'train'
 output_dir: './output'  # 当前不支持自定义修改，请勿修改该默认值
 load_checkpoint: ''
+src_strategy_path_or_dir: ''
 auto_trans_ckpt: False  # If true, auto transform load_checkpoint to load in distributed model
 only_save_strategy: False
 resume_training: False

configs/bert/run_bert_base_uncased.yaml

@@ -2,6 +2,7 @@ seed: 0
 run_mode: 'train'
 output_dir: './output'  # 当前不支持自定义修改，请勿修改该默认值
 load_checkpoint: ''
+src_strategy_path_or_dir: ''
 auto_trans_ckpt: False  # If true, auto transform load_checkpoint to load in distributed model
 only_save_strategy: False
 resume_training: False

configs/bert/run_bert_tiny_uncased.yaml

@@ -2,6 +2,7 @@ seed: 0
 run_mode: 'train'
 output_dir: './output'  # 当前不支持自定义修改，请勿修改该默认值
 load_checkpoint: ''
+src_strategy_path_or_dir: ''
 auto_trans_ckpt: False  # If true, auto transform load_checkpoint to load in distributed model
 only_save_strategy: False
 resume_training: False

configs/bloom/run_bloom_176b.yaml

@@ -2,6 +2,7 @@ seed: 0
 run_mode: 'train'
 output_dir: './output'  # 当前不支持自定义修改，请勿修改该默认值
 load_checkpoint: ''
+src_strategy_path_or_dir: ''
 auto_trans_ckpt: False  # If true, auto transform load_checkpoint to load in distributed model
 only_save_strategy: False
 resume_training: False

configs/bloom/run_bloom_560m.yaml

@@ -2,6 +2,7 @@ seed: 0
 run_mode: 'train'
 output_dir: './output'  # 当前不支持自定义修改，请勿修改该默认值
 load_checkpoint: ''
+src_strategy_path_or_dir: ''
 auto_trans_ckpt: False  # If true, auto transform load_checkpoint to load in distributed model
 only_save_strategy: False
 resume_training: False

configs/bloom/run_bloom_65b.yaml

@@ -2,6 +2,7 @@ seed: 0
 run_mode: 'train'
 output_dir: './output'  # 当前不支持自定义修改，请勿修改该默认值
 load_checkpoint: ''
+src_strategy_path_or_dir: ''
 auto_trans_ckpt: False  # If true, auto transform load_checkpoint to load in distributed model
 only_save_strategy: False
 resume_training: False

configs/bloom/run_bloom_7.1b.yaml

@@ -2,6 +2,7 @@ seed: 0
 run_mode: 'train'
 output_dir: './output'  # 当前不支持自定义修改，请勿修改该默认值
 load_checkpoint: ''
+src_strategy_path_or_dir: ''
 auto_trans_ckpt: False  # If true, auto transform load_checkpoint to load in distributed model
 only_save_strategy: False
 resume_training: False