Struct ReferenceVisionInput 

pub struct ReferenceVisionInput {
    pub patches: ReferenceTensor,
    pub positions: Vec<[u32; 2]>,
    pub output_tokens: usize,
}

Fields

patches: ReferenceTensor

Patch rows, row-major [patches, patch_row_width]. The pixel contract is per tower program:

• VisionPlan::Factored (gemma-4): raw pixels in [0, 1], width 3 * patch_size^2; the executor applies the graph’s 2x - 1.
• VisionPlan::Glm5Fused: PREPROCESSED pixels (rescale 1/255 then CLIP mean/std normalize, done by the image processor), width in_channels * temporal_patch_size * patch_size^2 in (c, t, ph, pw) flat order, token sequence in spatial-merge block-major order.

positions: Vec<[u32; 2]>

Per-patch 2D position. Factored: [x, y]. Glm5Fused: [h, w] (the upstream get_vision_position_ids column order), block-major over merge blocks.

output_tokens: usize

Trait Implementations

impl Clone for ReferenceVisionInput

fn clone(&self) -> ReferenceVisionInput

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ReferenceVisionInput

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl PartialEq for ReferenceVisionInput

fn eq(&self, other: &ReferenceVisionInput) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for ReferenceVisionInput