gdn_decode_wyn_table

Function gdn_decode_wyn_table 

Source
pub fn gdn_decode_wyn_table(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    h_tables: DevicePtr,
    query: DevicePtr,
    key: DevicePtr,
    value: DevicePtr,
    gate: DevicePtr,
    beta: DevicePtr,
    output: DevicePtr,
    hi_tables: DevicePtr,
    slab_entry_stride: u32,
    batch_size: u32,
    num_k_heads: u32,
    num_v_heads: u32,
    k_dim: u32,
    v_dim: u32,
    qk_stride: u32,
    v_stride: u32,
    gb_stride: u32,
    stream: u64,
) -> Result<()>
Expand description

2026-09-25: Cross-sequence batched-verify launch of the wyN pointer-table twins (gated_delta_rule_wy{5..16}_table and their _f16_table forms). state_is_table is compiled into the symbol, so the argument list is the same as gdn_decode_wyn, but the state arguments are tables:

  • h_tables: the layer’s staged table slice; slab 0 holds one per-sequence H base pointer per entry (VERIFY_WY_TABLE_SEQS entries).
  • hi_tables: slab 1 (Hi0); the following Hi slabs are VERIFY_WY_TABLE_SEQS entries apart.
  • slab_entry_stride must be VERIFY_WY_TABLE_SEQS as u32: the kernel reads the contiguous form’s stride argument as the number of pointer entries between Hi slabs.

The tables are staged by upload_verify_wy_tables (model-engine verify_e2.rs).