| { | |
| "model_type": "llama", | |
| "architectures": [ | |
| "DeepForCausalLM" | |
| ], | |
| "version": "0.13.0", | |
| "vocab_size": 32000, | |
| "hidden_size": 2048, | |
| "intermediate_size": 5632, | |
| "num_hidden_layers": 24, | |
| "num_attention_heads": 16, | |
| "num_key_value_heads": 8, | |
| "max_position_embeddings": 2048, | |
| "rope_theta": 10000.0, | |
| "rms_norm_eps": 1e-06, | |
| "attention_dropout": 0.0, | |
| "hidden_act": "silu", | |
| "tie_word_embeddings": true, | |
| "initializer_range": 0.02, | |
| "pad_token_id": 1, | |
| "bos_token_id": 2, | |
| "eos_token_id": 0, | |
| "use_token_routed_mlp": true, | |
| "num_experts": 4, | |
| "use_qk_norm": true, | |
| "use_sdpa": true, | |
| "sliding_window": null, | |
| "dynamics_alpha": 0.9, | |
| "dynamics_beta": 0.1, | |
| "dynamics_gate": 0.5, | |
| "dynamics_dt": 0.1, | |
| "dynamics_controller_hidden": 64, | |
| "parameters": "~1.5B", | |
| "innovations": [ | |
| "KQV Order (v0.13.0) - industry standard like Qwen, Llama, GPT", | |
| "Mu-Guided KQV (INL 2025) - mu biases K, Q, AND V", | |
| "Mu-Guided Expert Routing - mu influences MLP expert selection", | |
| "Mu Residual Highway - accumulated context across layers", | |
| "Token-Routed MLP with mu override", | |
| "INL Dynamics with velocity tracking", | |
| "Grouped Query Attention (GQA)", | |
| "RoPE positional embeddings", | |
| "QK Normalization", | |
| "Flash Attention (SDPA)" | |
| ] | |
| } |