Struct Packed 

pub struct Packed {
    pub cfg: Qwen4ExpConfig,
    pub manifest: Manifest,
    pub dir: PathBuf,
    pub dense: Mmap,
    pub experts: Mmap,
    pub experts_file: File,
    pub ngram: Mmap,
}

Fields

cfg: Qwen4ExpConfig

manifest: Manifest

dir: PathBuf

dense: Mmap

experts: Mmap

experts_file: File

Open handle on experts.bin for advisory reads (prefetch).

ngram: Mmap

Implementations

impl Packed

pub fn open(model_dir: &Path) -> Result<Self>

model_dir holds config.json and tokenizer.json; the packed files live in model_dir/packed, or directly in a standalone packed directory.

pub fn dense_bytes(&self, name: &str) -> Result<&[u8]>

pub fn shape(&self, name: &str) -> Result<&[usize]>

pub fn bf16(&self, name: &str) -> Result<&[bf16]>

pub fn i64s(&self, name: &str) -> Result<Vec<i64>>

pub(super) fn ngram_metadata(&self, prefix: &str) -> Result<NgramMetadata>

Read a PLE embedding’s hash parameters from the dense store for either backend.

pub fn qlinear(&self, prefix: &str) -> Result<QLinear<'_>>

Affine 4-bit group-64 linear layer {prefix}.{weight,scales,biases}.

pub fn expert_layer(&self, layer: usize) -> usize

Record index for a main decoder layer’s expert block.

pub fn mtp_expert_layer(&self, i: usize) -> Result<usize>

Record index for MTP layer i’s expert block.

pub fn record_offset(&self, record_layer: usize, expert: usize) -> usize

pub fn expert(&self, record_layer: usize, expert: usize) -> ExpertRef<'_>

pub fn ngram_row(&self, id: u64, dst: &mut [f32])

Dequantize one hashed n-gram row using the manifest’s width and group size.