nx_embedding.nx

nx_embedding.nx -- token-ID -> embedding-vector lookup. Closes the INPUT half of the inference pipeline. With this brick shipped, the substrate can do: token_ids -> nx_embedding_lookup -> [n_tok, hidden_dim] tensor (the input to layer 0) Composes only existing primitives: NxTensor (L1 container) NxQuantBlock / Q8 / Q4_K (L2 quantized weight containers) LoopVerdict (control) ===== Embedding table layout ==================================== An embedding table is a [vocab_size, hidden_dim] matrix. Token ID i selects row i. Two storage paths: Dense: *NxTensor with shape [vocab_size, hidden_dim], i64 Q10 values. Simple memcpy of row i to out[t, :]. Quantized: *NxQuantBlock (q4_0) or *NxQuantBlockQ8 (q8_0) or *NxQuantQ4K (q4_K). Row-by-row layout: vocab_size blocks of (hidden_dim / BLOCK_SIZE) sub-blocks each. Lookup dequantizes the requested row to Q10. VRAM impact: storing a 50k x 4096 embedding table dense in Q10 is 50000 * 4096 * 8 = 1.6 GB. Quantized at q4_0 is ~150 MB (10.67x compression). Per nx_quant_policy, embeddings stay at f16 by default (small + sensitive), but for memory-constrained deployment a quantized embedding table is a real ~1.5 GB save. ===== Out-of-vocab handling ===================================== Returns NX_EMB_ERR_OOV if any token ID >= vocab_size. Caller MUST validate token IDs upstream (or accept the error verdict and fall back). No silent UNK substitution -- explicit verdict. genealogy_id: mikolov_2013_word2vec + bengio_2003_neural_lm + rumelhart_hinton_williams_1986_distributed_repr