For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Python class

VisionInputs

`VisionInputs`

class max.pipelines.lib.role_contracts.VisionInputs(*, embeddings, scatter_indices)

source

Bases: Generic[_LeafT]

The image embeddings a DECODER merges into its token stream.

The framework passes them when a VISION module is registered. On a batch without new images both are empty.

Parameters:

embeddings: _LeafT

source

The encoded images, shape [num_image_tokens, hidden_size], replicated on every device of the decoder’s mesh.

scatter_indices: _LeafT

source

Each embedding row’s position in the decoder’s tokens, shape [num_image_tokens], replicated like embeddings.