add doc

Author: pan-x-c

docs/sphinx_doc/source/tutorial/trinity_configs.md

@@ -475,6 +475,7 @@ trainer:
   use_dynamic_bsz: true
   max_token_len_per_gpu: 16384
   ulysses_sequence_parallel_size: 1
+  max_checkpoints_to_keep: 5
   trainer_config: null
 ```
 
@@ -499,6 +500,7 @@ trainer:
 - `use_dynamic_bsz`: Whether to use dynamic batch size.
 - `max_token_len_per_gpu`:  The maximum number of tokens to be processed in forward and backward when updating the policy. Effective when `use_dynamic_bsz=true`.
 - `ulysses_sequence_parallel_size`: Sequence parallel size.
+- `max_checkpoints_to_keep`: Maximum number of checkpoints to keep. Older checkpoints will be deleted. If not specified, all checkpoints will be kept.
 - `trainer_config`: The trainer configuration provided inline.
 ---
 

docs/sphinx_doc/source_zh/tutorial/trinity_configs.md

@@ -472,6 +472,7 @@ trainer:
   use_dynamic_bsz: true
   max_token_len_per_gpu: 16384
   ulysses_sequence_parallel_size: 1
+  max_checkpoints_to_keep: 5
   trainer_config: null
 ```
 
@@ -496,6 +497,7 @@ trainer:
 - `use_dynamic_bsz`: 是否使用动态批量大小。
 - `max_token_len_per_gpu`: 训练过程中，每个 GPU 最大 token 长度; 当 `use_dynamic_bsz=true` 时生效。
 - `ulysses_sequence_parallel_size`: 序列并行的并行度，即用于分割单个序列的 GPU 数量。
+- `max_checkpoints_to_keep`: 保留的最大检查点数量。超过此数量后，最旧的检查点将被删除。
 - `trainer_config`: 内联提供的 trainer 配置。
 
 ---