bridge.recipes.nemotronh.gb200.nemotron_3_super

GB200 pretraining recipe for Nemotron 3 Super.

Module Contents

Functions

nemotron_3_super_pretrain_64gpu_gb200_bf16_config

Return the Nemotron 3 Super BF16 pretraining config for 64 GB200 GPUs.

Data

__all__

API

bridge.recipes.nemotronh.gb200.nemotron_3_super.nemotron_3_super_pretrain_64gpu_gb200_bf16_config() → megatron.bridge.training.config.ConfigContainer

Return the Nemotron 3 Super BF16 pretraining config for 64 GB200 GPUs.

This is the convergence-oriented counterpart of the canonical 64-GPU performance recipe. It uses the same GB200 parallel layout, HybridEP dispatcher, CUDA graph scopes, and process environment while retaining natural expert routing, runtime validation, and checkpointing.

Returns:

GB200 BF16 pretraining configuration.

bridge.recipes.nemotronh.gb200.nemotron_3_super.__all__

[‘nemotron_3_super_pretrain_64gpu_gb200_bf16_config’]