Struct GlmVit 

pub struct GlmVit {
    pub out_hidden_size: usize,
    /* private fields */
}

Fields

out_hidden_size: usize

Implementations

impl GlmVit

pub fn forward_batched( &self, images: &[(&[f32], usize, usize)], gpu: &dyn GpuBackend, stream: u64, ) -> Result<Vec<(usize, usize, usize)>>

Encode N images. Returns per-image (post_h, post_w, merged_p) in image order, matching the Qwen encoder’s contract so the prefill splice is shared.

Images are encoded in greedy chunks whose Σpatches fits p_max; the merged rows of every chunk land contiguously in buf_out at the global image order, so a batch that overflows the scratch degrades to more launches rather than to a wrong answer or an out-of-bounds write.

impl GlmVit

pub fn new( patch_embed_w: DevicePtr, patch_embed_b: DevicePtr, blocks: Vec<GlmVitBlock>, merger: GlmVitMerger, geo: &GlmVitGeometry, gpu: &dyn GpuBackend, ) -> Result<Self>

pub fn out_row(&self, row: usize) -> DevicePtr

Device row of the packed merged output — what the embedding splice copies over an <|image|> position.