katgpt-types

Shared configuration, RNG, math utilities, SIMD kernels, and inference types for the katgpt-rs / riir-engine inference framework. Pure substrate leaf — no katgpt-* dependencies.

Overview

The foundational leaf crate that every other katgpt-* crate (and riir-engine) builds on. Contains the shared Config struct (~1.5k lines), XorShift64 PRNG, SIMD-accelerated linear-algebra kernels (softmax / rmsnorm / matmul / ternary), CPU-side LoRA adapter types, domain embeddings, and inference result types.

The types and simd modules are co-located because types::math calls simd kernels and simd::ternary uses types::TernaryWeights — they form a tight bidirectional leaf that cannot be split further without breaking the cycle.

Key types / modules

Config / InferenceOverrides / kv_dim — the central inference configuration
Rng — XorShift64 PRNG
simd — NEON / AVX2 / WASM-SIMD128 / scalar fallback kernels
math — softmax, rmsnorm, matmul, sample_token, gegelu, silu, swiglu
lora — LoraAdapter, LoraPair, lora_apply
merkle — MerkleOctree, MerkleProof (BLAKE3 commitment)
kv_cache — QuantizedKVCache trait (shared extension point)
leaky_core — leaky_step (shared delta-rule primitive)
slod — ScaleBoundary (spectral LOD routing)
temporal — TemporalDerivativeKernel, sigmoid_surprise_gate
sense — SenseModule, ShardEmbedding, TernaryDir, DilationConfig

Feature flags

default = []. The following are tracking flags — they gate re-export visibility (mirroring the historical feature surface in katgpt-core), not structural code. All core substrate compiles unconditionally.

Feature	Description
`domain_latent`	DomainLatent embedding type (Plan 038)
`gpart_adapter`	GPart Isometric Partition adapter (Research 227)
`data_gate`	Task-level gating for self-play stability (Plan 111)
`sr2am_configurator`	SR²AM Configurator context types (Plan 112)
`hydra_budget`	Hydra adaptive layer budget types (Plan 165)
`deltanet_inference`	DeltaNet layer type enum (Plan 182)
`collapse_aware_thinking`	Collapse-aware ThinkingBudget enum (Plan 212)
`wall_attention`	Wall Attention WallConfig (Plan 173)
`sparse_mlp`	Sparse MLP matmul kernel (Plan 022)
`plasma_path`	Bit-plane ternary weights (Plan 148)
`maxsim`	MaxSim late-interaction scoring (Plan 080)
`sigmoid_margin`	Sigmoid margin loss (implies `maxsim`)
`rim_slots`	RiM Reasoning Buffer Slots (Plan 172)
`belief_drafter`	NextLat Belief-State Speculative Drafter (Plan 217)
`sia_feedback`	SIA FeedbackBandit enum variants (Plan 163)
`spectral_threat`	LinOSS Modal Threat Prediction fields (Plan 241)
`gpart_pruning`	GPart pruning stats (Plan 257)
`depth_invariance`	Depth-Invariance Diagnostic (Plan 306)

Dependencies

fastrand, blake3, serde, half, rayon

License