nemo_automodel.components.models.glm5_next.image_processing

Image-only backport of the GLM-5.3 dynamic patch processor.

Module Contents

Classes

Name	Description
`Glm5NextImageProcessor`	Dynamically resize, normalize and flatten GLM-5.3 image patches.
`Glm5NextImageProcessorKwargs`	Additional dynamic resize and patchification options.

Functions

Name	Description
`smart_resize`	Return an aligned H/W canvas inside the configured token budget.

Data

__all__

API

class nemo_automodel.components.models.glm5_next.image_processing.Glm5NextImageProcessor()

Bases: TorchvisionBackend

Dynamically resize, normalize and flatten GLM-5.3 image patches.

max_image_tokens

= 8000

merge_size

= 2

min_image_tokens

= 16

model_input_names

= ['pixel_values', 'image_grid_thw']

patch_expand_factor

= 1

patch_size

= 14

resample

= PILImageResampling.BICUBIC

rescale_factor

= 1 / 255

size

= {'longest_edge': 1}

temporal_patch_size

= 2

nemo_automodel.components.models.glm5_next.image_processing.Glm5NextImageProcessor._preprocess(
    images: list[torch.Tensor],
    do_resize: bool,
    size: transformers.image_utils.SizeDict,
    resample,
    do_rescale: bool,
    rescale_factor: float,
    do_normalize: bool,
    image_mean,
    image_std,
    patch_size: int,
    temporal_patch_size: int,
    merge_size: int,
    patch_expand_factor: int,
    min_image_tokens: int,
    max_image_tokens: int,
    disable_grouping: bool | None,
    return_tensors,
    kwargs = {}
) -> transformers.image_processing_utils.BatchFeature

Implement TorchvisionBackend’s grouped preprocessing contract.

nemo_automodel.components.models.glm5_next.image_processing.Glm5NextImageProcessor.get_number_of_image_patches(
    height: int,
    width: int,
    images_kwargs: dict | None = None
) -> int

Return the number of unmerged vision patches for one source image.

nemo_automodel.components.models.glm5_next.image_processing.Glm5NextImageProcessor.patchify(
    images: torch.Tensor,
    patch_size: int,
    merge_size: int,
    temporal_patch_size: int
) -> tuple[torch.Tensor, int, int]

staticmethod

Flatten block-major duplicated temporal patches.

nemo_automodel.components.models.glm5_next.image_processing.Glm5NextImageProcessor.resize(
    images: torch.Tensor,
    resample,
    factor: int,
    temporal_factor: int,
    min_image_tokens: int,
    max_image_tokens: int,
    kwargs = {}
) -> torch.Tensor

Aspect-preserving resize followed by right/bottom zero padding.

class nemo_automodel.components.models.glm5_next.image_processing.Glm5NextImageProcessorKwargs()

Bases: ImagesKwargs

Additional dynamic resize and patchification options.