Expand description
Recurrent-state cache contracts for state-space / hybrid models.
Recurrent state is intentionally separate from KV cache. KV grows with sequence length and is addressed through attention blocks; recurrent state is compact per-layer state that can coexist with KV in hybrid models.
Structs§
- Recurrent
State Handle Stats - Statistics for one recurrent-state handle.
- Recurrent
State Manager Stats - Aggregate recurrent-state manager statistics.
- Recurrent
State Spec - Allocation request for recurrent state.
- Recurrent
State Tensor Spec - A single recurrent-state tensor owned by a model layer.
Enums§
- Recurrent
State Resume Policy - Resume policy for state that has been preempted or evicted.
Traits§
- Recurrent
State Handle - Recurrent-state cache handle.
- Recurrent
State Manager - Recurrent-state cache manager for allocation and lifecycle management.