TINYLOOP_KV_INT8 | 1 | INT8 K/V with per-head scales. −49% KV footprint. |
TINYLOOP_KV_H_MODE | 1 | Store-h cache: reconstruct K/V from stored hidden state. Bit-exact. 50% saving. |
TINYLOOP_KV_INT8 + TINYLOOP_KV_H_MODE | 1 + 1 | h-as-INT8. ~66% saving. |
TINYLOOP_KV_INT4 + TINYLOOP_KV_H_MODE | 1 + 1 | h-as-INT4. 78% saving. Recommended for long-context. |
TINYLOOP_KV_H_FUSED | 1 | Fused INT4+unpack kernel. Currently performance-negative. |
TINYLOOP_PAGED_KV | 1 | Paged KV cache. 16-token pages, pool-based. ~5.6% overhead. Foundation for continuous batching. |