nx_f32_llm.nx

nx_f32_llm.nx -- full Llama-class model forward pass. The API a real text-in -> logits-out inference loop calls. Composes: embed lookup (token_id -> embedding row copy) nx_f32_llama_stack (N-layer transformer forward) nx_f32_rmsnorm (final pre-LM-head normalization) nx_f32_matmul (LM head: hidden -> logits) Algorithm: 1. Embedding lookup: for each token t in 0..n_tokens: x[t, :] = embed_weights[token_ids[t], :] 2. Stack forward through N layers (advances cache by n_tokens). 3. Final RMSNorm per token with gamma_out. 4. Logits = matmul(out_norm, lm_head) -> [n_tokens, vocab_size] The caller reads logits[(n_tokens-1)*vocab_size:n_tokens*vocab_size] to get next-token logits for autoregressive sampling. Earlier rows are valid for any per-position decode/classification. Tied-embed optimization: real Qwen/Llama variants often use the same matrix for embedding and LM head. Caller can pass the transposed embed_weights as lm_head -- this kernel does not care. genealogy_id: standard_llama_forward_pass lineage_id: substrate_f32_llm_v1