mirror of https://github.com/hpcaitech/ColossalAI
26b7aac0be
* [zero] refactor low-level zero folder structure * [zero] fix legacy zero import path * [zero] fix legacy zero import path * [zero] remove useless import * [zero] refactor gemini folder structure * [zero] refactor gemini folder structure * [zero] refactor legacy zero import path * [zero] refactor gemini folder structure * [zero] refactor gemini folder structure * [zero] refactor gemini folder structure * [zero] refactor legacy zero import path * [zero] fix test import path * [zero] fix test * [zero] fix circular import * [zero] update import |
||
|---|---|---|
| .. | ||
| model | ||
| utils | ||
| README.md | ||
| arguments.py | ||
| bert_dataset_provider.py | ||
| evaluation.py | ||
| hostfile | ||
| loss.py | ||
| nvidia_bert_dataset_provider.py | ||
| pretrain_utils.py | ||
| run_pretrain.sh | ||
| run_pretrain_resume.sh | ||
| run_pretraining.py | ||
README.md
Pretraining
- Pretraining roberta through running the script below. Detailed parameter descriptions can be found in the arguments.py.
data_path_prefixis absolute path specifies output of preprocessing. You have to modify the hostfile according to your cluster.
bash run_pretrain.sh
--hostfile: servers' host name from /etc/hosts--include: servers which will be used--nproc_per_node: number of process(GPU) from each server--data_path_prefix: absolute location of train data, e.g., /h5/0.h5--eval_data_path_prefix: absolute location of eval data--tokenizer_path: tokenizer path contains huggingface tokenizer.json, e.g./tokenizer/tokenizer.json--bert_config: config.json which represent model--mlm: model type of backbone, bert or deberta_v2
- if resume training from earylier checkpoint, run the script below.
bash run_pretrain_resume.sh
--resume_train: whether to resume training--load_pretrain_model: absolute path which contains model checkpoint--load_optimizer_lr: absolute path which contains optimizer checkpoint