nemo_curator.models.nemotron_h_vl

Module Contents

Classes

Name	Description
`LLM`	-
`NemotronHVL`	NemotronH hybrid Mamba-Attention VLM for video captioning.
`SamplingParams`	-

Data

EXPECTED_VIDEO_TAG_PARTS

NemotronVariant

VIDEO_TAG_SPLIT_MAX

VLLM_AVAILABLE

_NEMOTRON_REVISION_INFO

_NEMOTRON_VARIANTS_INFO

API

class nemo_curator.models.nemotron_h_vl.LLM()

class nemo_curator.models.nemotron_h_vl.NemotronHVL(
    model_dir: str,
    model_variant: nemo_curator.models.nemotron_h_vl.NemotronVariant = 'nemotron',
    caption_batch_size: int = 8,
    max_output_tokens: int = 512,
    stage2_prompt_text: str | None = None,
    verbose: bool = False
)

Bases: ModelInterface

NemotronH hybrid Mamba-Attention VLM for video captioning.

Supports multiple checkpoint variants from HuggingFace:

nemotron / nemotron-bf16: BF16 precision (default)
nemotron-fp8: FP8 quantized
nemotron-nvfp4: NVFP4 quantized

Models are automatically downloaded from HuggingFace on first use.

_hf_model_id

= _NEMOTRON_VARIANTS_INFO[self._normalized_variant]

_normalized_variant

NemotronVariant = 'nemotron-bf16'

model_id_names

list[str]

Return HuggingFace model ID for the selected variant.

stage2_prompt

weight_file

= str(Path(model_dir) / self._hf_model_id)

nemo_curator.models.nemotron_h_vl.NemotronHVL._refine_caption_prompt(
    original_prompt: str,
    refinement_text: str
) -> str

Create a refined prompt for stage 2 captioning.

nemo_curator.models.nemotron_h_vl.NemotronHVL.download_weights_on_node(
    model_dir: str,
    variant: nemo_curator.models.nemotron_h_vl.NemotronVariant = 'nemotron'
) -> None

classmethod

Download NemotronH VL weights from HuggingFace.

Models are automatically downloaded from HuggingFace Hub on first use. Supports multiple quantization variants for different performance/memory tradeoffs.

Parameters:

model_dir

str

Base directory for model weights. The model will be downloaded to a subdirectory named after the HuggingFace model ID.

variant

NemotronVariantDefaults to 'nemotron'

Model variant to download. Options:

“nemotron” or “nemotron-bf16”: BF16 precision (default)
“nemotron-fp8”: FP8 quantized
“nemotron-nvfp4”: NVFP4 quantized

nemo_curator.models.nemotron_h_vl.NemotronHVL.generate(
    videos: list[dict[str, typing.Any]],
    generate_stage2_caption: bool = False,
    batch_size: int = 16
) -> list[str]

nemo_curator.models.nemotron_h_vl.NemotronHVL.setup() -> None

class nemo_curator.models.nemotron_h_vl.SamplingParams()

nemo_curator.models.nemotron_h_vl.EXPECTED_VIDEO_TAG_PARTS = 2

nemo_curator.models.nemotron_h_vl.NemotronVariant = Literal['nemotron', 'nemotron-bf16', 'nemotron-fp8', 'nemotron-nvfp4']

nemo_curator.models.nemotron_h_vl.VIDEO_TAG_SPLIT_MAX = 1

nemo_curator.models.nemotron_h_vl.VLLM_AVAILABLE = True

nemo_curator.models.nemotron_h_vl._NEMOTRON_REVISION_INFO: Final = {'nemotron': '5d250e2e111dc5e1434131bdf3d590c27a878ade', 'nemotron-bf16': '5d250...

nemo_curator.models.nemotron_h_vl._NEMOTRON_VARIANTS_INFO: Final = {'nemotron': 'nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16', 'nemotron-bf16': 'nvi...