From 67f742fde2ba562316f45253effa7d817f4b73c2 Mon Sep 17 00:00:00 2001 From: Haixin Nan Date: Fri, 8 May 2026 00:41:37 +0800 Subject: [PATCH] Fix typos in training comments --- megatron/training/training.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/megatron/training/training.py b/megatron/training/training.py index 4c2b04d2618..793afb1dfb3 100644 --- a/megatron/training/training.py +++ b/megatron/training/training.py @@ -898,7 +898,7 @@ def pretrain( ft_integration.setup() timestamp_after_in_job_setup = time.time() - # Initalize and get arguments, timers, and Tensorboard writer. + # Initialize and get arguments, timers, and Tensorboard writer. initialize_megatron( get_embedding_ranks=get_embedding_ranks, get_position_embedding_ranks=get_position_embedding_ranks, @@ -1960,7 +1960,7 @@ def _save_state_dict(attr_name, label): if save_params_in_this_iteration: _save_state_dict(attr_name="data", label="params") - # when freezing sub-models we may have a mixture of successful and unsucessful ranks, + # when freezing sub-models we may have a mixture of successful and unsuccessful ranks, # so we must gather across mp ranks update_successful = logical_and_across_model_parallel_group(update_successful) # grad_norm and num_zeros_in_grad will be None on ranks without trainable params,