nx_transformer_stack.nx

nx_transformer_stack.nx -- multi-layer transformer forward composer. L5 brick. The missing layer between "load weights for one block" (nx_gguf_load_block) and "run the full LLM pipeline" (nx_llm_run v2 -- next brick). Composes: - Pre-allocated scratch tensors (one set, reused across layers) - Per-layer loop that: * loads block weights from GGUF via nx_gguf_load_block_weights * calls nx_transformer_block_forward with the loaded bundle - Optional final RMSNorm (output_norm.weight from load_model) Does NOT include: embed, output_projection, sample. Those compose in nx_llm_run v2. This brick proves multi-layer chain works. Bits-up composition (every primitive already canonical): nx_tensor.nx -- NxTensor + alloc nx_rmsnorm.nx -- final norm nx_transformer_block.nx -- per-layer forward + weight bundle nx_gguf.nx + nx_gguf_load.nx -- header + per-tensor load nx_gguf_load_block.nx -- 9-tensor per-layer loader genealogy_id: standard_transformer_stack_pattern + touvron_2023_llama_decoder lineage_id: substrate_transformer_stack_v1_single_head