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

VisionMerge

VisionMerge

class max.nn.graph_groups.VisionMerge(*args, **kwargs)

source

Bases: GraphGroup[str, tuple[TensorType, …] | tuple[tuple[TensorType, …], …]]

Image embeddings per device and the token positions they replace.

A model that injects image features into later layers, such as Qwen3-VL’s deepstack, holds one per-device tuple per layer.

Parameters:

• args (object)
• kwargs (object)

Return type:

Self

kind

kind: ClassVar[str | None] = 'vision_merge'

source

Identifies the feature wherever the group sits in a tree.