nx_model_spec.nx

nx_model_spec.nx -- typed model configuration envelope. L4 foundational data brick. Every model-running primitive composes against this: nx_gguf loader populates an NxModelSpec from metadata; the inference scaffold reads it to size scratch buffers and dispatch the right norm + activation + position-encoding kernels. Per the bits-up cardinal: model-specific constants (n_heads, hidden_dim, etc.) belong in a TYPED struct, not scattered as magic numbers across call sites. Per the data-driven cardinal: the same substrate runs Llama / Mistral / Qwen / GPT-2 / BERT just by swapping NxModelSpec values. ===== Sealed enums ============================================== Norm kind: NX_NORM_RMS -- Zhang/Sennrich 2019 (Llama / Mistral / Qwen) NX_NORM_LAYER -- Ba/Kiros/Hinton 2016 (GPT-2 / BERT / T5 / ViT) NX_NORM_NONE -- bypass (debug / experimentation) Activation kind (FFN gate function): NX_ACT_SILU -- Ramachandran 2017 (Llama / Mistral SwiGLU) NX_ACT_GELU -- Hendrycks/Gimpel 2016 (GPT-2 / BERT GeGLU) NX_ACT_RELU -- Fukushima 1969 / Glorot 2011 (older models) Position encoding kind: NX_POS_ROPE -- Su 2021 (Llama / Mistral / Qwen) NX_POS_SINUSOIDAL -- Vaswani 2017 (original transformer) NX_POS_LEARNED -- BERT / GPT-2 absolute-learned table NX_POS_NONE -- bypass FFN gating kind: NX_FFN_SWIGLU -- SiLU(gate) * up (Llama / Mistral) NX_FFN_GEGLU -- GELU(gate) * up (newer encoder models) NX_FFN_PLAIN -- act(up) (original transformer) genealogy_id: yaml_huggingface_config_2022 + onnx_model_metadata + ggml_gguf_metadata_2024 lineage_id: substrate_model_spec_v1