-
Notifications
You must be signed in to change notification settings - Fork 0
Home
Sulkysubject37 edited this page Jan 21, 2026
·
7 revisions
Purpose: This document defines the project's core mission, boundaries, and intended audience to prevent misuse.
VECTORIA is a deterministic, inspectable, and high-performance computational kernel framework. It serves as the "engine room" for higher-level applications that require absolute control over numerical execution, memory layout, and auditability.
Unlike general-purpose machine learning frameworks, VECTORIA prioritizes correctness over peak throughput and transparency over abstraction.
- Strict Determinism: Execution results are bitwise identical across repeated runs on the same hardware.
- Semantic Inference Stack: Includes stable implementations of LayerNorm, Softmax, CrossEntropy, Attention, Multi-Head Attention, and the full Transformer Encoder Block, all validated against reference math.
- Semantic Truth: The C++ Reference implementation defines the mathematical truth. SIMD kernels (ARM64 NEON, x86_64 AVX2) are validated against this truth.
- Explicit Control: Memory is managed via a static arena. There is no garbage collection, no hidden threads, and no dynamic graph mutation during execution.
- Auditable: Every decision—from kernel dispatch to memory allocation—is traceable.
- NOT an ML Framework: It has no auto-differentiation (autograd), no optimizers, and no training loops.
- NOT a Black Box: It contains no heuristics that swap algorithms at runtime based on opaque criteria.
- NOT "Fastest at All Costs": We will disable SIMD optimizations if they drift from the reference implementation's numerical results.
- Systems Engineers: Building verified compute pipelines.
- Researchers: Requiring reproducible numerical experiments.
- Developers: Needing a bridge between research code and constrained deployment environments (CoreML).