From c9c682d19c9e92a261c3879322016e61095e37cb Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E8=83=A1=E8=AF=91=E6=96=87?= <1020030101@qq.com> Date: Fri, 10 Jan 2025 01:50:51 +0800 Subject: [PATCH] [doc] deepspeed universal checkpoint (#35015) * universal checkpoint * Update docs/source/en/deepspeed.md Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com> * Update docs/source/en/deepspeed.md Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com> * Update docs/source/en/deepspeed.md Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com> --------- Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com> --- docs/source/en/deepspeed.md | 14 ++++++++++++++ 1 file changed, 14 insertions(+) diff --git a/docs/source/en/deepspeed.md b/docs/source/en/deepspeed.md index 7f7995c466..ad3d4240f8 100644 --- a/docs/source/en/deepspeed.md +++ b/docs/source/en/deepspeed.md @@ -586,6 +586,20 @@ You can choose the communication data type by setting the `communication_data_ty } ``` +### Universal Checkpointing + +[Universal Checkpointing](https://www.deepspeed.ai/tutorials/universal-checkpointing) is an efficient and flexible feature for saving and loading model checkpoints. It enables seamless model training continuation and fine-tuning across different model architectures, parallelism techniques, and training configurations. + +Resume training with a universal checkpoint by setting [load_universal](https://www.deepspeed.ai/docs/config-json/#checkpoint-options) to `true` in the config file. + +```yaml +{ + "checkpoint": { + "load_universal": true + } +} +``` + ## Deployment DeepSpeed can be deployed by different launchers such as [torchrun](https://pytorch.org/docs/stable/elastic/run.html), the `deepspeed` launcher, or [Accelerate](https://huggingface.co/docs/accelerate/basic_tutorials/launch#using-accelerate-launch). To deploy, add `--deepspeed ds_config.json` to the [`Trainer`] command line. It’s recommended to use DeepSpeed’s [`add_config_arguments`](https://deepspeed.readthedocs.io/en/latest/initialize.html#argument-parsing) utility to add any necessary command line arguments to your code.