Function apply_expert_lora_sorted 

pub fn apply_expert_lora_sorted(
    gpu: &dyn GpuBackend,
    kernels: &LoraKernels,
    layer: &ExpertLoraLayer,
    proj: ExpertProj,
    expert_offsets_host: &[u32],
    x: DevicePtr,
    base_out: DevicePtr,
    max_rows: u32,
    lora_xa: DevicePtr,
    lora_delta: DevicePtr,
    stream: u64,
) -> Result<()>

Fold proj’s per-expert LoRA deltas onto the SORTED grouped-GEMM output.

x = the projection’s sorted input ([total_expanded, pair.k_in]), base_out = the projection’s sorted output ([total_expanded, pair.n_out], modified in place). expert_offsets_host is the D2H copy of the device expert_offsets ([num_experts + 1]). One apply_lora_delta(m = rows) per adapted expert, over that expert’s contiguous row block — byte-identical to rows sequential m=1 folds. Only experts with a pair for proj AND non-zero routed rows launch.