bridge.recipes.nemotron_vl.nemotron_nano_v2_vl

Module Contents

Functions

nemotron_nano_v2_vl_12b_pretrain_config	Create a pre-training configuration for Nemotron Nano V2 VL.
nemotron_nano_v2_vl_12b_finetune_config	Create a finetuning configuration for Nemotron Nano V2 VL.

API

bridge.recipes.nemotron_vl.nemotron_nano_v2_vl.nemotron_nano_v2_vl_12b_pretrain_config(

dir: Optional[str] = None,

name: str = 'nemotron_nano_v2_vl_pretrain',

hf_model_path: str = 'nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16',

dataset_type: Optional[str] = None,

mock: bool = False,

dataset_maker_name: str = 'make_cord_v2_dataset',

tensor_parallelism: int = 4,

pipeline_parallelism: int = 1,

pipeline_parallelism_dtype: Optional[torch.dtype] = None,

virtual_pipeline_parallelism: Optional[int] = None,

context_parallelism: int = 1,

sequence_parallelism: bool = False,

train_iters: int = 300000,

global_batch_size: int = 32,

micro_batch_size: int = 2,

seq_length: int = 4096,

lr: float = 0.0003,

min_lr: float = 3e-05,

lr_warmup_iters: int = 500,

lr_decay_iters: Optional[int] = None,

precision_config: Optional[Union[megatron.bridge.training.mixed_precision.MixedPrecisionConfig, str]] = 'bf16_mixed',

comm_overlap_config: Optional[megatron.bridge.training.comm_overlap.CommOverlapConfig] = None,

save_interval: Optional[int] = 200,

) → megatron.bridge.training.config.ConfigContainer

Create a pre-training configuration for Nemotron Nano V2 VL.

Note: Current dataset pipeline is text-centric. To train multimodal tokens, your preprocessed data should include placeholder tokens (e.g., ) as needed.

bridge.recipes.nemotron_vl.nemotron_nano_v2_vl.nemotron_nano_v2_vl_12b_finetune_config(

*,

pretrained_checkpoint: str = '',

lora_on_language_model: bool = False,

lora_on_vision_model: bool = False,

save_checkpoint_dir: Optional[str] = None,

**pretrain_kwargs,

) → megatron.bridge.training.config.ConfigContainer

Create a finetuning configuration for Nemotron Nano V2 VL.

This helper wraps :func:nemotron_nano_v2_vl_12b_pretrain_config, forwarding all keyword arguments to it while additionally wiring up the :class:CheckpointConfig for finetuning from a given pretrained_checkpoint.

Parameters: pretrained_checkpoint: str Path to a Megatron-Bridge checkpoint (or a directory produced by convert_ckpt_hf_to_megatron) that will be loaded before training. save_checkpoint_dir: str | None, default run_output_dir / "checkpoints" Directory where new checkpoints will be saved / resumed from. If not provided, we reuse the default path chosen by nemotron_nano_v2_vl_12b_pretrain_config. lora_on_language_model: bool = True Whether to apply PEFT to the language model. lora_on_vision_model: bool = True Whether to apply PEFT to the vision model. **pretrain_kwargs: Any Additional keyword arguments are forwarded verbatim to :func:nemotron_nano_v2_vl_12b_pretrain_config to customise the base recipe (e.g. batch size, learning rate, parallelism).