Function load_moe_qwen35 

pub fn load_moe_qwen35(
    store: &WeightStore,
    layer_prefix: &str,
    num_experts: usize,
    gpu: &dyn GpuBackend,
    config: &ModelConfig,
    variant: Nvfp4Variant,
    absmax_k: KernelHandle,
    quantize_k: KernelHandle,
    stream: u64,
    copies: Nvfp4MoeCopies,
) -> Result<MoeWeights>

2026-09-25: Load a Qwen3.5 MoE block ({layer_prefix}.mlp) as NVFP4.

2026-09-27: copies names the projections to build; routed experts that this rank does not hold are ExpertWeight::null(). The fused layout builds every projection (it errors on anything but Nvfp4MoeCopies::ALL).