bridge.recipes.nemotronh.h100.nemotron_3_super

Module Contents

Functions

_configure_super_source	Choose checkpoint architecture before applying the existing recipe policy.
nemotron_3_super_pretrain_16gpu_h100_bf16_config	Return a pre-training config for Nemotron 3 Super (120B-A12B LatentMoE).
_apply_nemotron_3_super_64gpu_h100_training_stack	Apply the 64-H100 Nemotron 3 Super execution and batch configuration.
_nemotron_3_super_pretrain_64gpu_h100_bf16_config	Return the convergence-oriented Nemotron 3 Super config for 64 H100 GPUs.
nemotron_3_super_sft_16gpu_h100_bf16_config	Return a full SFT config for Nemotron 3 Super (120B-A12B LatentMoE).
nemotron_3_super_sft_16gpu_h100_bf16_32k_config	Return the 16-H100 BF16 32K full-SFT configuration.
nemotron_3_super_peft_1gpu_h100_bf16_config	Return a PEFT config for Nemotron 3 Super (120B-A12B LatentMoE).
nemotron_3_super_peft_16gpu_h100_bf16_config	Return the 16-H100 BF16 PEFT configuration.

Data

NEMOTRON_3_SUPER_HF_MODEL_ID
__all__

API

bridge.recipes.nemotronh.h100.nemotron_3_super.NEMOTRON_3_SUPER_HF_MODEL_ID

‘nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16’

bridge.recipes.nemotronh.h100.nemotron_3_super._configure_super_source(

cfg: megatron.bridge.training.config.ConfigContainer,

*,

hf_path: str | None,

text_only: bool,

revision: str | None,

trust_remote_code: bool,

) → None

Choose checkpoint architecture before applying the existing recipe policy.

bridge.recipes.nemotronh.h100.nemotron_3_super.nemotron_3_super_pretrain_16gpu_h100_bf16_config(

*,

hf_path: str | None = None,

text_only: bool = False,

revision: str | None = None,

trust_remote_code: bool = False,

) → megatron.bridge.training.config.ConfigContainer

Return a pre-training config for Nemotron 3 Super (120B-A12B LatentMoE).

This is a Latent MoE model with Multi-Token Prediction (MTP). Default parallelism:

• TP=8, PP=1, EP=16, SP=True

Returns:

Pre-training configuration for Nemotron 3 Super.

Return type:

ConfigContainer

bridge.recipes.nemotronh.h100.nemotron_3_super._apply_nemotron_3_super_64gpu_h100_training_stack(

cfg: megatron.bridge.training.config.ConfigContainer,

) → megatron.bridge.training.config.ConfigContainer

Apply the 64-H100 Nemotron 3 Super execution and batch configuration.

This helper owns the language-stack settings that are reusable by the multimodal Super variant. It deliberately does not replace cfg.model or its model-native MTP configuration.

Parameters:

cfg – A Nemotron 3 Super-family training configuration.

Returns:

The updated training configuration.

bridge.recipes.nemotronh.h100.nemotron_3_super._nemotron_3_super_pretrain_64gpu_h100_bf16_config(

*,

hf_path: str | None = None,

text_only: bool = False,

revision: str | None = None,

trust_remote_code: bool = False,

) → megatron.bridge.training.config.ConfigContainer

Return the convergence-oriented Nemotron 3 Super config for 64 H100 GPUs.

The public hardware-agnostic nemotron_3_super_pretrain_config alias exposes this configuration. Its execution layout matches the canonical H100 performance recipe while it retains natural routing, runtime checks, and checkpointing for real-data training evidence.

Returns:

BF16 pretraining configuration for 64 H100 GPUs.

bridge.recipes.nemotronh.h100.nemotron_3_super.nemotron_3_super_sft_16gpu_h100_bf16_config(

*,

hf_path: str | None = None,

text_only: bool = False,

revision: str | None = None,

trust_remote_code: bool = False,

) → megatron.bridge.training.config.ConfigContainer

Return a full SFT config for Nemotron 3 Super (120B-A12B LatentMoE).

Default parallelism: TP=8, PP=1, EP=16, SP=True

Returns:

ConfigContainer with all settings pre-configured for Nemotron 3 Super SFT.

bridge.recipes.nemotronh.h100.nemotron_3_super.nemotron_3_super_sft_16gpu_h100_bf16_32k_config(

*,

hf_path: str | None = None,

text_only: bool = False,

revision: str | None = None,

trust_remote_code: bool = False,

) → megatron.bridge.training.config.ConfigContainer

Return the 16-H100 BF16 32K full-SFT configuration.

Long-context SFT uses a distinct PP8/CP2 layout rather than the default SFT topology.

bridge.recipes.nemotronh.h100.nemotron_3_super.nemotron_3_super_peft_1gpu_h100_bf16_config(

peft_scheme: str | megatron.bridge.peft.base.PEFT = 'lora',

*,

hf_path: str | None = None,

text_only: bool = False,

revision: str | None = None,

trust_remote_code: bool = False,

) → megatron.bridge.training.config.ConfigContainer

Return a PEFT config for Nemotron 3 Super (120B-A12B LatentMoE).

Default parallelism: TP=1, PP=1, EP=1, SP=True

Parameters:

peft_scheme – PEFT scheme - “lora”, “dora”, or a custom PEFT instance.

Returns:

ConfigContainer with all settings pre-configured for Nemotron 3 Super PEFT.

bridge.recipes.nemotronh.h100.nemotron_3_super.nemotron_3_super_peft_16gpu_h100_bf16_config(

peft_scheme: str | megatron.bridge.peft.base.PEFT = 'lora',

*,

hf_path: str | None = None,

text_only: bool = False,

revision: str | None = None,

trust_remote_code: bool = False,

) → megatron.bridge.training.config.ConfigContainer

Return the 16-H100 BF16 PEFT configuration.

bridge.recipes.nemotronh.h100.nemotron_3_super.__all__

[‘nemotron_3_super_peft_1gpu_h100_bf16_config’, ‘nemotron_3_super_peft_16gpu_h100_bf16_config’, ‘nem…