Struct QuantizedLinear4BitPlannerEvidence 

pub struct QuantizedLinear4BitPlannerEvidence {

    pub in_dim: u32,
    pub out_dim: u32,
    pub group_size: u32,
    pub group_count: u32,
    pub packed_weight_bytes: u64,
    pub dequantized_weight_bytes: u64,
    pub sidecar_bytes: u64,
    pub bias_bytes: u64,
    pub output_bytes: u64,
    pub dequant_bytes_elided: u64,
    pub matmul_m: u32,
    pub matmul_k: u32,
    pub matmul_n: u32,
    pub matmul_tile: u32,
    pub matmul_selected_path: &'static str,
    pub matmul_candidate_path: Option<&'static str>,
    pub matmul_fallback_reason: Option<&'static str>,
    pub tensor_core_eligible: bool,
    pub output_drift_abs_tolerance: f32,
}

Planner evidence for fused grouped INT4 linear versus dequantized matmul.

Fields

in_dim: u32

Input feature dimension.

out_dim: u32

Output feature dimension.

group_size: u32

Quantization group size.

group_count: u32

Number of quantization groups.

packed_weight_bytes: u64

Packed INT4 weight bytes.

dequantized_weight_bytes: u64

Bytes that a materialized f32 dequantized weight matrix would require.

sidecar_bytes: u64

Scale plus zero-point sidecar bytes.

bias_bytes: u64

Bias bytes.

output_bytes: u64

Output bytes.

dequant_bytes_elided: u64

Dequantized weight bytes avoided by the fused path.

matmul_m: u32

Equivalent matmul planner M dimension.

matmul_k: u32

Equivalent matmul planner K dimension.

matmul_n: u32

Equivalent matmul planner N dimension.

matmul_tile: u32

Equivalent matmul planner K tile.

matmul_selected_path: &'static str

Selected shared matmul planner path.

matmul_candidate_path: Option<&'static str>

Candidate tensor-core path from the shared matmul planner, when any.

matmul_fallback_reason: Option<&'static str>

Shared matmul planner fallback reason, when the selected path is cooperative.

tensor_core_eligible: bool

Whether the shared matmul planner selected a tensor-core path.

output_drift_abs_tolerance: f32

Maximum absolute output drift accepted by evidence tests.

Trait Implementations

impl Clone for QuantizedLinear4BitPlannerEvidence

fn clone(&self) -> QuantizedLinear4BitPlannerEvidence

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for QuantizedLinear4BitPlannerEvidence

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

Formats the value using the given formatter.

impl PartialEq for QuantizedLinear4BitPlannerEvidence

fn eq(&self, other: &QuantizedLinear4BitPlannerEvidence) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for QuantizedLinear4BitPlannerEvidence