Struct SmoothQuantLinearArgs 

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

Args bundle for a SmoothQuant linear launch.

act_scale_scratch is a caller-owned [M] FP scratch buffer used to broadcast the descriptor’s per-tensor act_scale into the per-row form the underlying quantized_linear_w8a8 kernel consumes. Caller-owned so it can be reused across launches without re-allocation — the Plan’s workspace_size() returns 0.

Fields

act_q: TensorRef<'a, S8, 2>

Pre-quantized int8 activation [M, K].

weight_q: TensorRef<'a, TWQ, 2>

Pre-smoothed-then-quantized int8 weight [N, K].

weight_scale: TensorRef<'a, TIn, 1>

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

output: TensorMut<'a, TIn, 2>

FP output [M, N].

act_scale_scratch: TensorMut<'a, TIn, 1>

Scratch for the per-row broadcast of act_scale. [M] FP. Caller-owned; reused across launches. Populated by the plan before the matmul launch.