Struct Nf4Weight 

pub struct Nf4Weight {
    pub w_packed: Tensor,
    pub absmax: Tensor,
    pub n: usize,
    pub k: usize,
    pub block_size: usize,
}

A bitsandbytes-style NF4-quantized weight, materialized as the two Tensor inputs that fuel_graph::NodeHandle::nf4_matmul expects, plus the cached (n, k, block_size) geometry for the matmul builder.

The two inputs are kept as separate Tensor fields rather than bundled into a single tensor — that matches the fused op’s 3-input signature exactly. Self::matmul is the convenience builder for the common case where the caller has an activations: Tensor and just wants the linear-layer output.

Fields

w_packed: Tensor

Packed weight bytes, shape [N, K/2] U8.

absmax: Tensor

Per-block absmax scales, shape [N, K/block_size] F32.

n: usize

Output features (rows of the weight matrix).

k: usize

Input features (the K in (M, K) @ (K, N) → (M, N)).

block_size: usize

Quantization block size (typically 64 in bitsandbytes).

Implementations

impl Nf4Weight

pub fn matmul(&self, activations: &Tensor) -> Tensor

Convenience: run activations @ dequant(weight) via the fused fuel_graph::NodeHandle::nf4_matmul op. activations must live on the same graph as self.w_packed and self.absmax; the caller threads this by building activations from self.w_packed.graph_tensor() (or any other tensor on the same graph).

Output dtype matches activations’ dtype (F32 / F16 / BF16 — see fuel_graph::NodeHandle::nf4_matmul for the full contract).