Struct OpParams 

pub struct OpParams {
    pub weight_ptr: u64,
    pub gamma_ptr: u64,
    pub scalar: f32,
    pub int_param: u32,
    pub weight_qtype: u32,
    pub bias_ptr: u64,
    pub bias_len: usize,
}

Operation-specific parameters.

Fields

weight_ptr: u64

Weight pointer (for MulMat: quantized weights on device)

gamma_ptr: u64

Normalization gamma pointer (for RmsNorm)

scalar: f32

Scalar parameter (epsilon for RmsNorm, etc.)

int_param: u32

Integer parameter (position for RoPE, etc.)

weight_qtype: u32

PERF-050 (aprender#2753): GGML quantization type code of weight_ptr.

Without this the graph carried only a pointer, and the dispatcher had no way to know what it pointed at – so it assumed Q4_K for every MulMat. In a q4_k_m model that is right for 170 tensors and wrong for 29: attn_v, ffn_down and the LM head are Q6_K, and dequantizing Q6_K with the Q4_K kernel produced garbage of magnitude ~5e8 with NaNs.

The numeric GGML code is used rather than a Rust enum because this crate must not depend on the serving crate’s WeightQuantType. 12 = Q4_K, which is also the Default, so nodes that do not set it behave exactly as before.

bias_ptr: u64

PERF-050 (aprender#2769): bias vector applied after a MulMat, 0 when there is none.

The graph had no bias node and no way to express one, so the QKV bias Qwen2.5 models carry was silently dropped on this path while the M=1 path applied it. Defaults to 0, which means “no bias” and reproduces the previous behaviour exactly.

bias_len: usize

Length of bias_ptr in elements; 0 means no bias.

Trait Implementations

impl Clone for OpParams

fn clone(&self) -> OpParams

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for OpParams

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for OpParams

fn default() -> Self

Returns the “default value” for a type.