nx_vram_budget.nx

nx_vram_budget.nx -- the team's VRAM FOOTPRINT model + reducer (operator: reduce VRAM while MEETING/improving functionality). Rule #21 (Resource Awareness) made executable: know the budget, account for EVERY component, and find the smallest config that still clears the quality floor. RESEARCHED breakdown (not hand-waved): VRAM = weights + KV cache + activations + overhead. weights: n_params * weight_bits / 8. Quantize -> 4x (Q4) at ~1.2% perplexity. KV cache: 2(K+V) * layers * kv_heads * head_dim * seq * batch * kv_bits / 8. GROWS with context and at long context can RIVAL the weights. Most quantizers IGNORE it (leave it fp16). Q8 KV = 2x smaller at <0.1% loss; GQA (kv_heads << query_heads) = up to 8x smaller. activations + overhead: transient buffers + fragmentation (add ~15%). The reducer answers "reduce VRAM while meeting functionality" = MINIMISE footprint s.t. the total quality loss stays within the Council floor (quality favoured). Sizes in MB (params in millions). license_tier: ORIGINAL Refs: KV cache formula (Lyceum/HF); KV quant Q8<0.1% loss (TechPlained).