Struct QwenImageEncoder 

pub struct QwenImageEncoder { /* private fields */ }

Qwen2.5-VL prompt encoder. The component CMF retains the official model.* and visual.* tensor names, plus U8 image.*_json assets.

Implementations

impl QwenImageEncoder

pub fn open(path: &Path) -> Result<Self, String>

Open a standalone Qwen2.5-VL component CMF. Sharded CMFs are opened through the core loader so the language and vision mappings remain lazy and no giant F32 embedding copy is created.

Examples found in repository

examples/qwen_image_encoder_fixture.rs (line 159)

fn main() -> Result<(), Box<dyn Error>> {
    let mut args = std::env::args_os().skip(1);
    let fixture_path = PathBuf::from(args.next().ok_or("missing fixture JSON")?);
    let cmf_path = PathBuf::from(args.next().ok_or("missing output CMF")?);
    let fixture: Fixture = serde_json::from_slice(&std::fs::read(&fixture_path)?)?;
    if fixture.grid != vec![1usize, 6, 10] || fixture.ids.len() <= 64 {
        return Err(format!(
            "unexpected fixture geometry grid={:?} ids={}",
            fixture.grid,
            fixture.ids.len()
        )
        .into());
    }
    if fixture.image_rgb.len() != fixture.image_width as usize * fixture.image_height as usize * 3 {
        return Err("fixture RGB byte count does not match image dimensions".into());
    }
    pack(&fixture, &cmf_path)?;
    let model = CmfModel::open(&cmf_path)?;
    if !model.verify().is_empty() {
        return Err(format!("fixture CMF failed verification: {:?}", model.verify()).into());
    }
    let encoder = QwenImageEncoder::open(&cmf_path)?;
    let image = image::RgbImage::from_raw(
        fixture.image_width,
        fixture.image_height,
        fixture.image_rgb.clone(),
    )
    .ok_or("fixture RGB image could not be constructed")?;
    let output = encoder.encode(&fixture.prompt, &[image])?;
    if output.hidden_size != 12 || output.seq_len * output.hidden_size != fixture.expected.len() {
        return Err(format!(
            "native conditioning shape {}×{} does not match oracle {} values",
            output.seq_len,
            output.hidden_size,
            fixture.expected.len()
        )
        .into());
    }
    check(&output.hidden, &fixture.expected)?;
    println!(
        "encoder tiny fixture PASS: tokens={} rows={} grid={:?}",
        fixture.ids.len(),
        output.seq_len,
        fixture.grid
    );
    Ok(())
}

pub fn hidden_size(&self) -> usize

pub fn model_uid(&self) -> u64

Stable CMF identity for the caller’s targeted GPU buffer release after this encoder’s stage scope ends. The encoder keeps this value available without exposing its internal model mapping.

pub fn encode( &self, prompt: &str, images: &[RgbImage], ) -> Result<Conditioning, String>

Encode a positive or negative prompt with the same reference image list. The caller invokes this once for each true-CFG branch; image preprocessing and vision placeholders therefore remain identical.

Examples found in repository

examples/qwen_image_encoder_fixture.rs (line 166)

fn main() -> Result<(), Box<dyn Error>> {
    let mut args = std::env::args_os().skip(1);
    let fixture_path = PathBuf::from(args.next().ok_or("missing fixture JSON")?);
    let cmf_path = PathBuf::from(args.next().ok_or("missing output CMF")?);
    let fixture: Fixture = serde_json::from_slice(&std::fs::read(&fixture_path)?)?;
    if fixture.grid != vec![1usize, 6, 10] || fixture.ids.len() <= 64 {
        return Err(format!(
            "unexpected fixture geometry grid={:?} ids={}",
            fixture.grid,
            fixture.ids.len()
        )
        .into());
    }
    if fixture.image_rgb.len() != fixture.image_width as usize * fixture.image_height as usize * 3 {
        return Err("fixture RGB byte count does not match image dimensions".into());
    }
    pack(&fixture, &cmf_path)?;
    let model = CmfModel::open(&cmf_path)?;
    if !model.verify().is_empty() {
        return Err(format!("fixture CMF failed verification: {:?}", model.verify()).into());
    }
    let encoder = QwenImageEncoder::open(&cmf_path)?;
    let image = image::RgbImage::from_raw(
        fixture.image_width,
        fixture.image_height,
        fixture.image_rgb.clone(),
    )
    .ok_or("fixture RGB image could not be constructed")?;
    let output = encoder.encode(&fixture.prompt, &[image])?;
    if output.hidden_size != 12 || output.seq_len * output.hidden_size != fixture.expected.len() {
        return Err(format!(
            "native conditioning shape {}×{} does not match oracle {} values",
            output.seq_len,
            output.hidden_size,
            fixture.expected.len()
        )
        .into());
    }
    check(&output.hidden, &fixture.expected)?;
    println!(
        "encoder tiny fixture PASS: tokens={} rows={} grid={:?}",
        fixture.ids.len(),
        output.seq_len,
        fixture.grid
    );
    Ok(())
}

pub fn encode_pair( &self, prompt: &str, negative_prompt: &str, images: &[RgbImage], ) -> Result<(Conditioning, Conditioning), String>

Name matching the pipeline’s positive/negative branch operation while keeping one shared image list and one immutable encoder mapping.