Struct QuantizedLinearArgs 

pub struct QuantizedLinearArgs<'a, TIn: Element, TWQ: IntElement> {
    pub activation: TensorRef<'a, TIn, 2>,
    pub weight_q: TensorRef<'a, TWQ, 2>,
    pub weight_scale: TensorRef<'a, TIn, 1>,
    pub output: TensorMut<'a, TIn, 2>,
    pub act_q_scratch: TensorMut<'a, S8, 2>,
    pub act_scale_scratch: TensorMut<'a, TIn, 1>,
}

Args bundle for a quantized_linear launch.

The caller supplies the already-quantized weight + its per-channel scale (offline-computed). The activation is FP; per-token activation quantization happens inside QuantizedLinearPlan::run via an internally orchestrated super::DynamicRangeQuantizePlan pass.

act_q_scratch and act_scale_scratch are caller-owned scratch buffers for the quantized activation + computed per-row activation scale. They are part of the args bundle (not workspace) so callers can reuse them across launches without re-allocation — the Plan’s workspace_size() returns 0.

Fields

activation: TensorRef<'a, TIn, 2>

FP activation [M, K].

weight_q: TensorRef<'a, TWQ, 2>

Already-quantized int8 weight [C_out, K].

weight_scale: TensorRef<'a, TIn, 1>

Per-output-channel weight scale [C_out] in FP.

output: TensorMut<'a, TIn, 2>

FP output [M, C_out].

act_q_scratch: TensorMut<'a, S8, 2>

Scratch for the per-token quantized activation [M, K] in int8. Caller-owned; reused across launches.

act_scale_scratch: TensorMut<'a, TIn, 1>

Scratch for the per-token activation scale [M] in FP. Caller-owned; reused across launches. Populated by the internally orchestrated dynamic-range pass.