nx_f32_llama_block_v4.nx

nx_f32_llama_block_v4.nx -- per-layer forward with lazy weights. Mirrors nx_f32_llama_block (v3) but accepts NxF32LazyWeight for each of the 7 matmul weights, dispatching to f32 or Q4_K backends per tensor. RMSNorm gammas remain as raw *i64 (they're 1D vectors, always small enough that lazy dequant isn't needed). When real GGUF files have most weights as Q4_K (e.g. Qwen2.5-0.5B- Instruct Q4_K_M), this forward avoids the eager-dequant memory blow-up: each matmul streams the Q4_K bytes through nx_q4k_to_f32 one row at a time, peak working memory drops from O(k*n) to O(n). genealogy_id: vaswani_2017 + touvron_2023_llama + ggml_format_taxonomy + tagged_dispatch lineage_id: substrate_f32_llama_block_v4_lazy