Struct Qwen3Encoder 

pub struct Qwen3Encoder {
    pub hidden: usize,
    /* private fields */
}

Fields

hidden: usize

Implementations

impl Qwen3Encoder

pub fn from_cmf(model: &Arc<CmfModel>) -> Result<Self, String>

pub fn set_exact_q8(&mut self, on: bool)

Weight-only exact q8 projections (f32 activations; see q8_exact_matmat). Off by default, so every existing caller keeps its numbers; Z-Image turns it on.

pub fn set_device_ops(&mut self, spec: &str)

Route the named projections (all, or a comma list of PROJ_NAMES) through the device GEMM of their codec; the host kernels stay the fallback. Experiment knob (CMF_ZIMAGE_TE_DEV in the Z-Image pipeline, ZC_TE_DEV in examples/zimage_techeck.rs).

pub fn encode(&self, ids: &[u32]) -> Vec<f32>

Causal full-sequence forward. Returns [n, hidden] — the residual stream leaving the last layer, normed only if the config says the checkpoint has a final norm (H3’s does not).

pub fn encode_with_images( &self, ids: &[u32], spans: &[ImageSpan], embeds: &[Vec<f32>], deepstack: &[Vec<f32>], ) -> Vec<f32>

Full forward with images. embeds replaces the token embedding at [span.start, span.start + span.len) with the vision tower’s merged tokens; deepstack[k] is added at the visual positions after LM layer k — the first layers, not the vision layers the features were taken from.

pub fn out_hidden(&self) -> usize

Width of what encode returns — the DiT’s conditioning width, which is the projection’s output when one is packed and the encoder’s own hidden size otherwise.