Skip to navigation

nemo_automodel.components.models.lfm3.processing

View as MarkdownOpen in Claude

LFM3-VL processor assembled from the Qwen image and video processors of the pinned Transformers release.

Module Contents

Classes

NameDescription
Lfm3VlProcessorQwen3-VL prompt construction over LFM3-VL’s tokenizer and patch settings.
Lfm3VlProcessorKwargsCall-time defaults: LFM3-VL takes no multimodal token-type ids.

API

class nemo_automodel.components.models.lfm3.processing.Lfm3VlProcessor()

Bases: Qwen3VLProcessor

Qwen3-VL prompt construction over LFM3-VL’s tokenizer and patch settings.

Released checkpoints name image and video processors that exist only as remote code. Transformers 5.23 maps LFM3-VL to its Qwen2-VL image and Qwen3-VL video processors instead; from_pretrained builds those from the checkpoint’s stored settings, so no remote code runs.

model_input_names
list[str]

Names of the tensors the processor returns, without Qwen3-VL’s multimodal token-type ids.

pretrained_model_name_or_path: str | os.PathLike,
kwargs: str | bool | None = {}
classmethod

Load the tokenizer, chat template and patch settings of an LFM3-VL checkpoint.

Parameters:

pretrained_model_name_or_path
str | os.PathLike

Hub repository id or local checkpoint directory.

**kwargs
str | bool | NoneDefaults to {}

Hub loading options such as revision, cache_dir and token.

Returns: Lfm3VlProcessor

Processor backed by the native Qwen image and video processors.

class nemo_automodel.components.models.lfm3.processing.Lfm3VlProcessorKwargs()

Bases: ProcessingKwargs

Call-time defaults: LFM3-VL takes no multimodal token-type ids.

_defaults