Expose hardcoded Megatron infrastructure parameters to user config#2230
Open
nic-nvidia wants to merge 2 commits intoNVIDIA-NeMo:mainfrom
Open
Expose hardcoded Megatron infrastructure parameters to user config#2230nic-nvidia wants to merge 2 commits intoNVIDIA-NeMo:mainfrom
nic-nvidia wants to merge 2 commits intoNVIDIA-NeMo:mainfrom
Conversation
Read checkpoint, timeout, and diagnostic settings from megatron_cfg with backward-compatible defaults instead of hardcoding them. New megatron_cfg fields (all optional, existing defaults preserved): - async_save, fully_parallel_save, fully_parallel_load, load_rng - distributed_timeout_minutes - logging_level New distributed_data_parallel_config field: - check_for_nan_in_grad Closes NVIDIA-NeMo#2229
3 tasks
yuki-97
reviewed
Apr 11, 2026
Contributor
There was a problem hiding this comment.
Hi @nic-nvidia , thanks for the enhancement! LGTM except the default config place. Could you help to update?
Also @yaoyu-33 @cuichenx , could you help to check whether the params in this PR are well supported in MBrdige?
| # Step 1: Setup distributed | ||
| setup_distributed() | ||
| setup_distributed( | ||
| timeout_minutes=config.get("megatron_cfg", {}).get("distributed_timeout_minutes"), |
Contributor
There was a problem hiding this comment.
We encourage to set default value in config.yaml instead of in code, so that people can know what feature we have and their default behavior w/o looking into the code.
Can you help to:
- Update to the below, also other configs
- Add the param (set to the default value) to several base configs? (other configs will inherit from the base one so don't need to change)
examples/configs/distillation_math.yamlexamples/configs/dpo.yamlexamples/configs/grpo_math_1B.yamlexamples/configs/rm.yamlexamples/configs/sft.yamlexamples/nemo_gym/grpo_nanov3.yamlexamples/nemo_gym/grpo_workplace_assistant_nemotron_nano_v2_9b.yamlresearch/template_project/configs/grpo_math_1B.yaml
Suggested change
| timeout_minutes=config.get("megatron_cfg", {}).get("distributed_timeout_minutes"), | |
| timeout_minutes=config["megatron_cfg"]["distributed_timeout_minutes"], |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
megatron_cfgwith backward-compatible defaults instead of hardcoding them insetup.pymegatron_cfgfields:async_save,fully_parallel_save,fully_parallel_load,load_rng,distributed_timeout_minutes,logging_leveldistributed_data_parallel_configfield:check_for_nan_in_gradCloses #2229
Test plan
test_basic_checkpoint_configpasses (backward compat, no config arg)test_checkpoint_config_overridesvalidates all 4 checkpoint fields