Struct GemmaVisionConfig 

pub struct GemmaVisionConfig {
    pub patch_size: usize,
    pub model_patch_size: usize,
    pub mm_embed_dim: usize,
    pub mm_posemb_size: usize,
    pub num_soft_tokens: usize,
    pub output_proj_dims: usize,
    pub pooling_kernel_size: usize,
    pub rms_norm_eps: f64,
}

Vision tower / projection config for Gemma 4 unified.

The HF nested vision_config carries these fields; defaults match google/gemma-4-12B.

Fields

patch_size: usize

Patch grid step in input pixels.

model_patch_size: usize

“Macro patch” used by Gemma 4 unified — model_patch_size pixels per soft-token-output tile.

mm_embed_dim: usize

Embedding dimension of the per-patch projection (before pool).

mm_posemb_size: usize

Number of positional embeddings (max patch grid the model ever sees).

num_soft_tokens: usize

Number of soft tokens the projector outputs for one image.

output_proj_dims: usize

Final projection target — matches the LM hidden_size so the soft tokens can be spliced into the text token stream.

pooling_kernel_size: usize

Square pooling kernel applied between patch projection and the soft-token down-sampler.

rms_norm_eps: f64

RMS norm epsilon used inside the projector.

Trait Implementations

impl Clone for GemmaVisionConfig

fn clone(&self) -> GemmaVisionConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for GemmaVisionConfig

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

Formats the value using the given formatter.

impl Default for GemmaVisionConfig

fn default() -> GemmaVisionConfig

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for GemmaVisionConfig

where GemmaVisionConfig: Default,

fn deserialize<__D>( __deserializer: __D, ) -> Result<GemmaVisionConfig, <__D as Deserializer<'de>>::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.