Struct DeviceRouteScratch 

pub struct DeviceRouteScratch<'a, B: Backend> {
    pub selected_ids: &'a mut B::Buffer,
    pub pair_weights: &'a mut B::Buffer,
    pub pairs_by_token: &'a mut B::Buffer,
    pub packed_token_idx: &'a mut B::Buffer,
    pub expert_offsets: &'a mut B::Buffer,
    pub sorted_tokens: &'a mut B::Buffer,
    pub block_ids: &'a mut B::Buffer,
    pub total_post_pad: &'a mut B::Buffer,
}

Bundle of pre-allocated device buffers for the graph-capturable device-routing path in moe_forward_bucketed. Pass Some to take the device path (under FERRUM_MOE_DEVICE_ROUTE=1); pass None for the legacy host-mediated path (used by tests + the non-vLLM CUDA bucketed path).

Pre-allocated on Qwen3MoeScratch (route_pairs_dev etc.) so the per-layer call doesn’t alloc inside a captured stream.

Fields

selected_ids: &'a mut B::Buffer

pair_weights: &'a mut B::Buffer

pairs_by_token: &'a mut B::Buffer

packed_token_idx: &'a mut B::Buffer

expert_offsets: &'a mut B::Buffer

sorted_tokens: &'a mut B::Buffer

block_ids: &'a mut B::Buffer

total_post_pad: &'a mut B::Buffer