| # Steps to run continued pretraining | |
| 1. Install the environment from as given in `multilinguality_megatron/Readme.md` | |
| 2. Run the following commands | |
| ```bash | |
| conda activate towerllm-env | |
| bash multilinguality_megatron/convert2megatron.sh | |
| bash multilinguality_megatron/model_sharding.sh | |
| bash multilinguality_megatron/continue_pretraining.sh | |
| ``` | |
| Arguments to take care of: | |
| ```bash | |
| convert2megatron.sh | |
| --megatron_model: Path where the megatron weights are to be saved | |
| --model: Path of huggingface model (KshitijAmbilduke/extended_non_uniform_model_tinyllama) | |
| --size: 1 (for TinyLlama) | |
| --repo: Location of the multilingual megatron repository | |
| model_sharding.sh | |
| --megatron_model: Path where the megatron weights are saved | |
| --sharded_model: Path of folder to save shards of the model | |
| --tp: Number of shards to create. (Number of shards == Number of GPUs used) | |
| --vocab_size: 37005 (32000+5005) | |
| continue_pretraining.sh | |
| --data_path="1 data/data_text_document" | |
| megatron_model: Path of folder containing sharded model | |
| model_dir: Path for folding storing checkpoints | |
| tokenizer_path: Path of extended tokenizer | |
| tp: number of shards | |
| ``` |