topic: intfp
21 modules sharing the intfp name family (derived from the tree's prefix discipline).
The 'intfp' topic family focuses on integer-based fixed-point training within the Nishi sovereign ecosystem, aiming to achieve efficient and precise computation without floating-point dependencies. Modules like nx_intfp_gradcheck_gate establish foundational verification for integer training, while nx_intfp_layer_gradcheck_gate integrates these principles into full transformer layers. nx_intfp_flashattn_gate enhances efficiency through optimized attention mechanisms, collectively supporting scalable, resource-conscious model development.
auto-narrated by the local model from this topic's module headers; links verified against the wiki index.
narrated overview -- maintained by the narration lane, module links verified against this wiki.
| module | description | lines | funcs |
|---|---|---|---|
| nx_intfp_attention_gradcheck_gate.nx | THE COMPOSITION PROOF: a full single-head CAUSAL self-attention block | 117 | 8 |
| nx_intfp_block_gradcheck_gate.nx | CAPSTONE assembly pattern: a pre-norm residual block y = x + Attn(RMSNorm(x)) | 165 | 9 |
| nx_intfp_dpo_gradcheck_gate.nx | DPO preference alignment (the RLHF/DPO axis) in Q20 INTEGER, gradchecked, no | 90 | 11 |
| nx_intfp_flashattn_gate.nx | FLASH-ATTENTION (online-softmax, O(T) memory, no T×T matrix) in Q20 INTEGER. | 67 | 8 |
| nx_intfp_frontier_census.nx | honest census: the sovereign INTEGER (no-float) training stack vs the 2026 LLM | 59 | 6 |
| nx_intfp_gradcheck_gate.nx | FIRST verified brick of FIXED-POINT INTEGER TRAINING (the sovereign fix for the | 80 | 5 |
| nx_intfp_layer_gradcheck_gate.nx | FULL integer transformer LAYER: two pre-norm residual blocks stacked | 176 | 16 |
| nx_intfp_lm_real_gate.nx | REAL-SCALE sovereign integer transformer LM. d=64, ffn=256 (4x), token+POSITION | 114 | 18 |
| nx_intfp_lm_simd_gate.nx | REAL-SCALE sovereign integer transformer LM. d=64, ffn=256 (4x), token+POSITION | 182 | 23 |
| nx_intfp_lm_train_gate.nx | COMPLETE integer transformer LM, trained on next-token prediction. Composes the | 130 | 18 |
| nx_intfp_log_gate.nx | integer natural LOG (fp_log) in Q20, completing the sovereign transcendental set | 54 | 5 |
| nx_intfp_lora_gradcheck_gate.nx | LoRA / parameter-efficient fine-tuning in Q20 INTEGER, gradchecked, no float. | 81 | 7 |
| nx_intfp_mla_gradcheck_gate.nx | 2026 FRONTIER: MLA (Multi-head Latent Attention, DeepSeek's marquee KV- | 98 | 8 |
| nx_intfp_mlp_gradcheck_gate.nx | SECOND brick of integer training: prove the CHAIN RULE in fixed-point. | 102 | 5 |
| nx_intfp_moe_gradcheck_gate.nx | 2026 FRONTIER capability in the sovereign integer stack: MIXTURE-OF-EXPERTS | 116 | 8 |
| nx_intfp_qknorm_gradcheck_gate.nx | 2026 FRONTIER: QK-NORM (Qwen3.5) in Q24 INTEGER, gradchecked, no float. | 101 | 9 |
| nx_intfp_rmsnorm_gradcheck_gate.nx | LAST transcendental for integer transformer training: RMSNorm needs | 92 | 6 |
| nx_intfp_rope_gate.nx | RoPE (Rotary Position Embedding, universal in 2026) in Q20 INTEGER, no float. Adds | 83 | 8 |
| nx_intfp_softmax_gradcheck_gate.nx | KEYSTONE op for integer ATTENTION training: fixed-point SOFTMAX with a | 85 | 6 |
| nx_intfp_swiglu_gradcheck_gate.nx | the OTHER transformer sublayer in integer: SwiGLU FFN, gradchecked. | 100 | 9 |
| nx_intfp_train_gate.nx | THIRD brick of integer training: prove an integer model actually LEARNS. | 111 | 4 |