nx_gguf_load_block.nx

nx_gguf_load_block.nx -- per-layer Llama-class weight bundle loader. L5 brick: composes nx_gguf_load_tensor + nx_rope_compute_inv_freq + NxTransformerBlockWeights to read the 9 named tensors for one transformer layer from a parsed GGUF and populate the bundle struct the transformer block forward consumes. This is the brick that takes nx_llm_run from v1 "API composition proof" to v2 "real-weight runner" -- caller can now load a Llama Q8_0 GGUF, ask for layer N's weights, and feed them to nx_transformer_block_forward unchanged. ===== Llama GGUF tensor naming convention (public format spec) === Per ggml docs (Gerganov 2024) -- clean-room implementation from public spec; no copied code. blk.{N}.attn_norm.weight -> gamma_attn (RMS-norm scale, 1-D) blk.{N}.attn_q.weight -> W_q blk.{N}.attn_k.weight -> W_k blk.{N}.attn_v.weight -> W_v blk.{N}.attn_output.weight -> W_o blk.{N}.ffn_norm.weight -> gamma_ffn (RMS-norm scale, 1-D) blk.{N}.ffn_gate.weight -> W_gate blk.{N}.ffn_up.weight -> W_up blk.{N}.ffn_down.weight -> W_down inv_freq is COMPUTED from rope_base via nx_rope_compute_inv_freq, not loaded -- GGUF stores rope_freq_base as metadata, not as a tensor. Caller supplies the base. Bits-up composition: nx_gguf.nx -- header + tensor_info table nx_gguf_load.nx -- per-tensor load + dequant (F32/F16/Q8_0) nx_strconv.nx -- decimal layer-index formatting nx_tensor.nx -- NxTensor return shape nx_rope.nx -- inv_freq computation nx_transformer_block.nx -- NxTransformerBlockWeights struct genealogy_id: llama_gguf_naming_gerganov_2024 +