Home

cortex-tensor

Pure-Rust tensor, transformer, and Mixture-of-Experts building blocks.

Version: 0.1.0 | Edition: 2024 | License: MIT OR Apache-2.0

cortex-tensor is a minimal, framework-free numerical kernel for transformer LMs and MoE routing. Extracted from corinth-canal and stripped of GPU/CUDA/Julia/SNN orchestration so it can stand alone as reusable open-source math.

Design Goals

• Zero GPU coupling — no cust, no pinned-host registration, no CUDA features
• Zero ML frameworks — no candle, tch, or ort; tensors are row-major Vec<f32>
• Small deps — serde, serde_json, thiserror, rand, rayon, memmap2, half (+ optional sentry)
• Inference-ready MoE — GGUF bridge with family-aware adapters (OLMoE, Qwen3-MoE, Gemma-4, DeepSeek-2, Llama-MoE)

Start Here

Page	Description
Getting Started	Install, matmul, transformer, router
Project Structure	Layout and modules
Architecture	Layers, ownership, non-goals
Tensor	Tensor type and element ops
Tensor Ops	matmul, norms, mask, embedding
Transformer	Attention, block, TransformerLM
MoE Router	OlmoeRouter, routing modes
GGUF Bridge	Checkpoint, dequant, quant types
Model Families	Adapters and synapse sources
Parser Boundary	engram-parser / safetensors plan
Errors	CortexError / HybridError
Sentry	Optional monitoring feature
CI and Quality	GHA, Azure, Codecov, Qodana, Snyk
Testing	Unit tests and env hooks
Ecosystem	Sibling crates
Glossary	Terms

Quick Commands

cargo build
cargo test
cargo fmt --all -- --check
cargo clippy --all-targets --all-features -- -D warnings

License

Dual-licensed under MIT or Apache-2.0 at your option.

---

Last updated: July 29, 2026 Updated by: Grok Build: Grok 4.5 Package tip reference: 37c5a21 (main, through PR #21)