Struct PagedDecodeTokenData 

pub struct PagedDecodeTokenData {
    pub token_ids: Arc<RwLock<Storage>>,
    pub rope_cos: Arc<RwLock<Storage>>,
    pub rope_sin: Arc<RwLock<Storage>>,
    pub block_table: Arc<RwLock<Storage>>,
    pub context_lens: Arc<RwLock<Storage>>,
    pub offset: Option<Arc<RwLock<Storage>>>,
}

Per-token re-bound data Consts for one PAGED decode step — the paged twin of DecodeTokenData. The shapes are stable across tokens (that is the whole point of the plan-once path); only the bytes change, so each is an Arc<RwLock<Storage>> overwriting the held PagedDecodeSession’s base-cache entry for its stable NodeId.

Fields

token_ids: Arc<RwLock<Storage>>

The single decode token id ([1] U32).

rope_cos: Arc<RwLock<Storage>>

RoPE cos table at THIS token’s absolute position ([1, head_dim] F32).

rope_sin: Arc<RwLock<Storage>>

RoPE sin table at THIS token’s absolute position ([1, head_dim] F32).

block_table: Arc<RwLock<Storage>>

Capacity-padded block table ([1, max_blocks_cap] U32) — the shape is FIXED by max_blocks_cap; the content grows a physical block whenever this token crosses a block boundary (padded tail stays 0, masked out).

context_lens: Arc<RwLock<Storage>>

Attended prefix length for this row ([1] U32) — the Op::PagedAttn mask cutoff (Ge(pos, context_len) → −inf) that neutralizes both the future positions AND the padded block_table tail.

offset: Option<Arc<RwLock<Storage>>>

The flattened KV-write offset (linear = phys·block_size + slot as a rank-0 I64), present ONLY on the device-offset path (CPU/CUDA, where Op::WriteSliceDoff reads the start device-side). None on the SymEnv path (Vulkan), where the offset rides write_sym in the per-pass env.