nx_llm_run.nx

nx_llm_run.nx -- end-to-end LLM inference scaffold. L5 integration brick. Closes the SUBSTRATE COMPLETENESS PROOF: every primitive shipped this session (BPE + embedding + RMSNorm + attention + RoPE + SiLU + sampler + ...) composes into a SINGLE callable that maps text -> next-token-text. Per the bits-up cardinal: no new math, no new kernels -- pure composition. ===== End-to-end pipeline ======================================= text input -> nx_bpe_encode (token_ids) -> nx_embedding_lookup or _q8 (input vectors) -> for each layer in 0..n_layers: nx_transformer_block_forward -> final RMSNorm (or LayerNorm) on output -> output projection (matmul to vocab_size) -> nx_attn_softmax_row_q10 (treat logits as 1-row tensor) -> nx_logit_apply_temperature -> nx_logit_top_k_mask -> nx_sample_categorical (next token) -> nx_bpe_decode (token -> text bytes) ===== Per the bits-up cardinal =================================== This file ONLY composes existing primitives. No new structs, no new sealed enums, no new math. Imports surface every layer: nx_tensor (NxTensor L1 container) nx_image (only via downstream consumers; not direct) nx_bpe (text <-> tokens) nx_embedding (token -> vector) nx_model_spec (config envelope) nx_transformer_block (per-layer compute) nx_rmsnorm / nx_layernorm (final norm) nx_attention (softmax + scale) nx_token_sample (logits -> token) nx_loop (bounded loops) ===== v1 scope ===================================================