bridge.recipes.nemotronh_multimodal.h100.nemotron_35_super_vl

Nemotron 3.5 Super VL training recipes.

Module Contents

Functions

_nemotron_35_super_vl_base	Create the model and training base owned by Nemotron 3.5 Super VL.
nemotron_35_super_vl_pretrain_64gpu_h100_bf16_config	Return the 64-H100 BF16 pretraining config for Super VL.
nemotron_35_super_vl_sft_64gpu_h100_bf16_config	Return the 64-H100 BF16 SFT configuration for Nemotron 3.5 Super VL.
nemotron_35_super_vl_peft_16gpu_h100_bf16_config	Return the 16-H100 BF16 PEFT configuration for Nemotron 3.5 Super VL.

Data

NEMOTRON_35_SUPER_VL_HF_MODEL_ID
NEMOTRON_35_SUPER_VL_HF_REVISION
__all__

API

bridge.recipes.nemotronh_multimodal.h100.nemotron_35_super_vl.NEMOTRON_35_SUPER_VL_HF_MODEL_ID

‘nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16’

bridge.recipes.nemotronh_multimodal.h100.nemotron_35_super_vl.NEMOTRON_35_SUPER_VL_HF_REVISION: str | None

None

bridge.recipes.nemotronh_multimodal.h100.nemotron_35_super_vl._nemotron_35_super_vl_base() → megatron.bridge.training.config.ConfigContainer

Create the model and training base owned by Nemotron 3.5 Super VL.

bridge.recipes.nemotronh_multimodal.h100.nemotron_35_super_vl.nemotron_35_super_vl_pretrain_64gpu_h100_bf16_config() → megatron.bridge.training.config.ConfigContainer

Return the 64-H100 BF16 pretraining config for Super VL.

The language stack starts from the Nemotron 3 Super H100 recipe, then uses the measured Super-VL pipeline balance, selective recompute, and HybridEP policy needed by the trainable vision stack. It retains natural expert routing, numerical guards, and checkpointing for real-data training. All present model stacks remain trainable.

Returns:

The Super-VL pretraining configuration.

bridge.recipes.nemotronh_multimodal.h100.nemotron_35_super_vl.nemotron_35_super_vl_sft_64gpu_h100_bf16_config() → megatron.bridge.training.config.ConfigContainer

Return the 64-H100 BF16 SFT configuration for Nemotron 3.5 Super VL.

The model and audio-free image/video Energon data path reuse the Nemotron Omni stack. The language decoder reuses the measured Nemotron 3 Super TP1/PP2/EP32 HybridEP training stack with two MTP prediction depths sharing one attention+MoE block and a separate temporal video embedder.

The Energon shard path must be set with dataset.path=<path>. Samples may contain text, images, and videos, but must not contain audio.

Returns:

The Super-VL SFT configuration.

bridge.recipes.nemotronh_multimodal.h100.nemotron_35_super_vl.nemotron_35_super_vl_peft_16gpu_h100_bf16_config(

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

) → megatron.bridge.training.config.ConfigContainer

Return the 16-H100 BF16 PEFT configuration for Nemotron 3.5 Super VL.

The base language, image, projector, and temporal-video weights remain frozen. Adapters target the model-native attention, Mamba, and MLP linear projections inherited from Nemotron 3 Super. The Energon dataset preserves Omni’s processor-sized image and temporal-video metadata contract. A two-stage pipeline fits the frozen base without activation recompute.

Parameters:

peft_scheme – PEFT scheme, either "lora", "dora", or a configured PEFT instance.

Returns:

The Super-VL PEFT configuration.

bridge.recipes.nemotronh_multimodal.h100.nemotron_35_super_vl.__all__

[‘NEMOTRON_35_SUPER_VL_HF_MODEL_ID’, ‘NEMOTRON_35_SUPER_VL_HF_REVISION’, ‘nemotron_35_super_vl_peft_…