bridge.recipes.nemotronh.gb200.nemotron_3_nano#
GB200 training recipes for Nemotron 3 Nano.
Module Contents#
Functions#
Return the Nemotron 3 Nano BF16 pretraining config for eight GB200 GPUs. |
|
Return the Nemotron 3.5 Lightning BF16 8K pretraining config. |
|
Return the Nemotron 3.5 Lightning BF16 8K Megatron FSDP config. |
|
|
Return the optimized TP1 4K packed OpenMathInstruct-2 SFT config. |
Data#
API#
- bridge.recipes.nemotronh.gb200.nemotron_3_nano._NEMOTRON_3_NANO_MODEL_ID#
‘nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16’
- bridge.recipes.nemotronh.gb200.nemotron_3_nano.nemotron_3_nano_pretrain_8gpu_gb200_bf16_config() megatron.bridge.training.config.ConfigContainer#
Return the Nemotron 3 Nano BF16 pretraining config for eight GB200 GPUs.
The recipe retains the established optimizer, scheduler, routing, and BF16 contracts. It applies the validated GB200 TP1/EP8 HybridEP topology and uses a 4,096-token sequence length for the paired NeMo-CI convergence workload.
- Returns:
GB200 BF16 pretraining configuration.
- bridge.recipes.nemotronh.gb200.nemotron_3_nano.nemotron_3_5_lightning_pretrain_8k_config() megatron.bridge.training.config.ConfigContainer#
Return the Nemotron 3.5 Lightning BF16 8K pretraining config.
- bridge.recipes.nemotronh.gb200.nemotron_3_nano.nemotron_3_5_lightning_pretrain_8k_fsdp_config() megatron.bridge.training.config.ConfigContainer#
Return the Nemotron 3.5 Lightning BF16 8K Megatron FSDP config.
- bridge.recipes.nemotronh.gb200.nemotron_3_nano.nemotron_3_5_lightning_sft_openmathinstruct2_packed_tp1_config() megatron.bridge.training.config.ConfigContainer#
Return the optimized TP1 4K packed OpenMathInstruct-2 SFT config.
- bridge.recipes.nemotronh.gb200.nemotron_3_nano.nemotron_3_nano_gb200_pretrain_config#
None
- bridge.recipes.nemotronh.gb200.nemotron_3_nano.__all__#
[‘nemotron_3_5_lightning_pretrain_8k_config’, ‘nemotron_3_5_lightning_pretrain_8k_fsdp_config’, ‘nem…