nx_llm_run_v2.nx

nx_llm_run_v2.nx -- end-to-end Llama-class inference composer (real weights edition). L6 brick. The CLOSER for the inference path: composes every loader + per-layer compute + sampler primitive shipped this session into a single callable that maps: text prompt + GGUF + BPE vocab + sampling knobs -> next-token-id Compared to nx_llm_run.nx (v1, scaffold-only): * v1 omitted the per-layer transformer compute (no weight loader) * v1 simulated output projection as identity copy * v2 sources ALL weights from a parsed GGUF and runs the full forward stack Per the bits-up + no-skipping cardinals: no new math, no reinvented kernels. Pure composition of: nx_bpe.nx_bpe_encode -- text -> token ids nx_gguf_load_model.nx_..._weights -- top-level tensor bundle nx_tensor.nx_t_alloc -- working buffers nx_embedding.nx_embedding_lookup -- token-id -> hidden vector nx_transformer_stack.nx_..._fwd -- N-layer forward nx_blas_i64.nx_blas_matmul -- output projection to vocab nx_token_sample.* + nx_attention.nx_attn_softmax_row_q10 -- temp + top_k + softmax + sample genealogy_id: standard_decoder_only_llm_inference + touvron_2023_llama + radford_2019_gpt2 lineage_id: substrate_llm_run_v2_real_weights