Struct StepVisionTower 

pub struct StepVisionTower { /* private fields */ }

Implementations

impl StepVisionTower

pub fn load(e: &Engine, dir: &Path) -> Result<Self, Box<dyn Error>>

Load the tower from the serving artifact’s own directory (the vision tensors live unquantized, BF16, inside the NVFP4 checkpoint: model.vision_model.* + model.vit_large_projector.weight, routed by model.safetensors.index.json). Refuses any directory whose tensors do not census as this exact program.

pub fn out_width(&self) -> usize

Embedding width this tower emits per row (the projector out_features == the trunk n_embd of the checkpoint it loaded from; 4096 on Step-3.7-Flash).

pub fn forward( &self, e: &Engine, patches: &[f32], g: usize, ) -> Result<CudaSlice<f32>, Box<dyn Error>>

Forward ONE view (the 728 main image at g=52, or one 504 crop tile at g=36): host patch rows [g*g, 588] -> device [rows, out_width] projector output, rows = (g/4 rounded per the two stride-2 convs)^2 (169 or 81).

pub fn forward_unit( &self, e: &Engine, unit: &StepVisionUnit, rows: &mut CudaSlice<f32>, row_off: usize, ) -> Result<usize, Box<dyn Error>>

Forward one whole unit (tiles first, then the main view — the vendor merge order) into rows at row_off * out_width. Returns rows written.