diff --git a/launcher_scripts/conf/training/mt5/170m.yaml b/launcher_scripts/conf/training/mt5/170m.yaml index 65726ad0c..1db7e5654 100755 --- a/launcher_scripts/conf/training/mt5/170m.yaml +++ b/launcher_scripts/conf/training/mt5/170m.yaml @@ -53,8 +53,8 @@ exp_manager: buffer_size: 5 model: - mcore_t5: False - transformer_engine: False + mcore_t5: True + transformer_engine: True # model parallelism micro_batch_size: 64 @@ -215,4 +215,4 @@ model: favor_longer_ngrams: False respect_document_boundaries: True # If true, a single training exampl cannot cross document boundaries, increasing the fraction of tokens within a batch. index_mapping_dir: null # path to save index mapping .npy files, by default will save in the same location as data_prefix - data_prefix: null # Should be weight path weight path... for a blended dataset. If null will automatically blend all language files in mC4_dir. + data_prefix: null # Should be weight path weight path... for a blended dataset. If null will automatically blend all language files in mC4_dir. \ No newline at end of file