`bridge.recipes.nemotronh.h100.nemotron_3_nano_4b`#

Recipes for the dense NVIDIA Nemotron 3 Nano 4B model.

Module Contents#

Functions#

`_model_config`	Build the exact dense Nemotron 3 Nano 4B architecture.
`_configure_tokenizer`	Pin the model-native tokenizer revision.
`_configure_kernels`	Set the common BF16 execution and memory policy.
`_configure_optimizer`	Apply the bounded model-card convergence cohort optimizer contract.
`nemotron_3_nano_4b_pretrain_8gpu_h100_bf16_config`	Return the bounded random-init pretraining config for eight H100 GPUs.
`nemotron_3_nano_4b_sft_8gpu_h100_bf16_config`	Return the 2K packed full-SFT config for eight H100 GPUs.
`nemotron_3_nano_4b_sft_8gpu_h100_bf16_32k_config`	Return the packed 32K full-SFT config with CP=2 for eight H100 GPUs.
`nemotron_3_nano_4b_peft_8gpu_h100_bf16_config`	Return the packed attention-LoRA config for eight H100 GPUs.

Data#

`_HF_MODEL_ID`
`_HF_MODEL_REVISION`
`__all__`

API#

bridge.recipes.nemotronh.h100.nemotron_3_nano_4b._HF_MODEL_ID#: ‘nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16’

bridge.recipes.nemotronh.h100.nemotron_3_nano_4b._HF_MODEL_REVISION#: ‘dfaf35de3e30f1867dd8dbc38a7fc9fb52d3914f’

bridge.recipes.nemotronh.h100.nemotron_3_nano_4b._model_config( *, seq_length: int, context_parallel_size: int = 1, ) → megatron.bridge.models.hybrid.hybrid_provider.HybridModelProvider#: Build the exact dense Nemotron 3 Nano 4B architecture.

bridge.recipes.nemotronh.h100.nemotron_3_nano_4b._configure_tokenizer( cfg: megatron.bridge.training.config.ConfigContainer, ) → None#: Pin the model-native tokenizer revision.

bridge.recipes.nemotronh.h100.nemotron_3_nano_4b._configure_kernels( cfg: megatron.bridge.training.config.ConfigContainer, ) → None#: Set the common BF16 execution and memory policy.

bridge.recipes.nemotronh.h100.nemotron_3_nano_4b._configure_optimizer( cfg: megatron.bridge.training.config.ConfigContainer, *, lr: float, min_lr: float, grad_reduce_in_fp32: bool, ) → None#: Apply the bounded model-card convergence cohort optimizer contract.

bridge.recipes.nemotronh.h100.nemotron_3_nano_4b.nemotron_3_nano_4b_pretrain_8gpu_h100_bf16_config() → megatron.bridge.training.config.ConfigContainer#

Return the bounded random-init pretraining config for eight H100 GPUs.

Recommended parallelism is TP=1, PP=1, CP=1, DP=8. The recipe is random initialization by default; loading the released HF checkpoint is not part of this recipe’s verification contract.

bridge.recipes.nemotronh.h100.nemotron_3_nano_4b.nemotron_3_nano_4b_sft_8gpu_h100_bf16_config() → megatron.bridge.training.config.ConfigContainer#

Return the 2K packed full-SFT config for eight H100 GPUs.

Recommended parallelism is TP=1, PP=1, CP=1, DP=8.

bridge.recipes.nemotronh.h100.nemotron_3_nano_4b.nemotron_3_nano_4b_sft_8gpu_h100_bf16_32k_config() → megatron.bridge.training.config.ConfigContainer#

Return the packed 32K full-SFT config with CP=2 for eight H100 GPUs.

Recommended parallelism is TP=1, PP=1, CP=2, DP=4.

bridge.recipes.nemotronh.h100.nemotron_3_nano_4b.nemotron_3_nano_4b_peft_8gpu_h100_bf16_config( peft_scheme: str | megatron.bridge.peft.base.PEFT = 'lora', ) → megatron.bridge.training.config.ConfigContainer#

Return the packed attention-LoRA config for eight H100 GPUs.

Parameters:: peft_scheme – PEFT scheme ("lora", "dora"), or a custom PEFT instance.