From 0500d6b0f6e049a3ceb6bd2813de95d9be8fb4d1 Mon Sep 17 00:00:00 2001 From: Anna Shors Date: Wed, 11 Dec 2024 10:49:24 -0800 Subject: [PATCH] link to mcore documentation (#11538) Signed-off-by: ashors1 --- nemo/lightning/pytorch/strategies/megatron_strategy.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/nemo/lightning/pytorch/strategies/megatron_strategy.py b/nemo/lightning/pytorch/strategies/megatron_strategy.py index f1a68f4225a6..8767df7dad03 100644 --- a/nemo/lightning/pytorch/strategies/megatron_strategy.py +++ b/nemo/lightning/pytorch/strategies/megatron_strategy.py @@ -158,7 +158,8 @@ class MegatronStrategy(DDPStrategy, io.IOMixin): Defaults to True. ckpt_load_strictness (StrictHandling, optional): defines loading strictness. If not None, overwrites the `strict` flag passed to `load_checkpoint`. - Defaults to None. + Defaults to None. For a list of supported values, refer to the Megatron Core documentation: + https://github.com/NVIDIA/Megatron-LM/blob/d4e72c0d33edc0c53aeb624f617eb77cebce6ae9/megatron/core/dist_checkpointing/validation.py#L46 setup_optimizers (bool): Whether to call the trainer's setup_optimizers function to perform any necessary conversions of optimizer parameters and move optimizer parameters to the correct device. Defaults to True.