pub fn mtp_attn_meta_dev(
base: DevicePtr,
max_blocks_per_seq: u32,
) -> AttnMetadataDevExpand description
2026-09-25: Pack one MTP attention-metadata slab, or return an error when
it would not fit region_bytes.
Layout: position u32 at 0, global_slot i64 at 8, seq_len i32 at 16,
then the block table from MTP_META_HEADER_BYTES, four little-endian
bytes per entry. region_bytes is what the caller owns at the
destination: the per-sequence stride for the batched propose, or
scratch_bytes - MTP_META_OFFSET for the single-sequence callers.
2026-09-29: The attention metadata pack_mtp_attn_meta lays out at base (position @0,
slot @8, sequence length @16, block table @MTP_META_HEADER_BYTES), for one row.