nx_lowrank_kv_budget.nx

nx_lowrank_kv_budget.nx -- the HARDWARE PAYOFF of data-driven low-rank KV (nx_lowrank_rank_plan) on the operator's real budget: RTX 5080, 16 GiB VRAM (rule #21 resource awareness; telos = capable AI on AFFORDABLE consumer hardware). KV cache is the context-length bottleneck: full KV grows O(n_layers * 2 * kv_dim) per token; data-driven low-rank KV grows O(n_layers * (r_K + r_V)) per token. This computes the MAX CONTEXT LENGTH full-vs-low-rank for a given model on a given VRAM budget. Pure integer math, NO slow forward. Real Qwen2.5-0.5B config + ranks (measured); 7B shown as an honest illustrative extrapolation. 100% sovereign (nx_cc->nxasm, no gcc). license_tier: ORIGINAL ⚠ SUPERSEDED 2026-06-17: the context-multipliers here (42x for 0.5B, 170x for 7B) used K=1/V=5 ranks from a DEGENERATE-context KV spectrum. The definitive seq=64 measurement (nx_lowrank_kv_real M4d) shows real KV needs avg rank K=27/V=37 of 64 for 90% energy => the real KV VRAM win is ~2-3x, NOT 42x. The budget MATH here is valid; the input ranks were wrong. WEIGHTS are heterogeneous too (attn low-rank but small, big ffn NOT low-rank) => weight-SVD also modest. BOTH low-rank VRAM levers are MODEST on real Qwen-0.5B.