bridge.recipes.nemotronh.gb200.nemotron_3_ultra#

GB200 functional pretraining recipe for Nemotron 3 Ultra.

Module Contents#

Functions#

_nemotron_3_ultra_pretrain_256gpu_gb200_bf16_ep16_config

Build the verified Nemotron 3 Ultra BF16 config for 256 GB200 GPUs.

nemotron_3_ultra_pretrain_256gpu_gb200_bf16_ep16_config

Return a verified Nemotron 3 Ultra BF16 pretrain config for 256 GB200 GPUs.

nemotron_3_ultra_pretrain_256gpu_gb200_bf16_ep16_deterministic_config

Return a deterministic Nemotron 3 Ultra BF16 pretrain config for 256 GB200 GPUs.

Data#

API#

bridge.recipes.nemotronh.gb200.nemotron_3_ultra._nemotron_3_ultra_pretrain_256gpu_gb200_bf16_ep16_config() → megatron.bridge.training.config.ConfigContainer#

Build the verified Nemotron 3 Ultra BF16 config for 256 GB200 GPUs.

bridge.recipes.nemotronh.gb200.nemotron_3_ultra.nemotron_3_ultra_pretrain_256gpu_gb200_bf16_ep16_config() → megatron.bridge.training.config.ConfigContainer#

Return a verified Nemotron 3 Ultra BF16 pretrain config for 256 GB200 GPUs.

The recipe uses TP4/PP4/EP16 natural-routing HybridEP with targeted expert activation offload and selective attention/layernorm recompute. It retains numerical checks and the library recipe’s optimizer, scheduler, and training objective.

Returns:

GB200 BF16 distributed-optimizer pretraining configuration.

bridge.recipes.nemotronh.gb200.nemotron_3_ultra.nemotron_3_ultra_pretrain_256gpu_gb200_bf16_ep16_deterministic_config() → megatron.bridge.training.config.ConfigContainer#

Return a deterministic Nemotron 3 Ultra BF16 pretrain config for 256 GB200 GPUs.

Wraps :func:nemotron_3_ultra_pretrain_256gpu_gb200_bf16_ep16_config and applies

Func:

~megatron.bridge.recipes.utils.determinism_utils.apply_determinism_overrides and spreads DETERMINISM_ENV_VARS into the recipe’s static process environment.

Returns:

GB200 BF16 deterministic pretraining configuration.

bridge.recipes.nemotronh.gb200.nemotron_3_ultra.__all__#

[‘nemotron_3_ultra_pretrain_256gpu_gb200_bf16_ep16_config’, ‘nemotron_3_ultra_pretrain_256gpu_gb200_…