Class QwenVisionTower

java.lang.Object
smile.deep.layer.LayerBlock
smile.llm.model.qwen.QwenVisionTower
All Implemented Interfaces:
Function<Tensor,Tensor>, Layer

public class QwenVisionTower extends LayerBlock
Qwen3.8 / Qwen3.5 vision tower (model.visual.*).

Patch embed is a bias linear equivalent to Conv3d (temporal, patch, patch) over flattened patches. DeepStack is rejected when indexes are non-empty.

  • Constructor Details

    • QwenVisionTower

      public QwenVisionTower(QwenVisionArgs args)
      Creates a vision tower for the given hyperparameters.
      Parameters:
      args - vision hyperparameters.
  • Method Details

    • args

      public QwenVisionArgs args()
      Returns the vision hyperparameters.
      Returns:
      vision hyperparameters.
    • forward

      public Tensor forward(Tensor input)
      Description copied from interface: Layer
      Forward propagation (or forward pass) through the layer.
      Parameters:
      input - the input tensor.
      Returns:
      the output tensor.
    • forward

      public Tensor forward(Tensor pixelValues, int[][] gridThw)
      Encodes packed patches to LLM vision tokens.
      Parameters:
      pixelValues - [N, patchDim].
      gridThw - [numMedia][3] = (T,H,W) patch units.
      Returns:
      [N/merge^2, outHidden] (caller owns).