Loom Documentation

M-POLY-VTD Engine Docs

Complete reference for Loom's poly package — the Multi-numerical POLYmorphic Volumetric Tiled-tensor Dispatcher that powers every Loom integration.

Architecture Overview Quick Reference GitHub

Where to start?

New? Read Overview → Layers → Training. Deploying to web or mobile? Go to Deployment. Need a snippet? Quick Reference has everything copy-paste ready.

Architecture Layers Training Deployment GPU Quick Reference

Architecture Overview

Big-picture architecture: the 3D grid, six design pillars, key types, and performance benchmarks.

Quick Reference

Concise copy-paste snippets for all common operations. Start here if you just need code.

Every layer type — Dense, CNN, RNN, MHA, SwiGLU, RMSNorm, Residual, Softmax, Parallel, Sequential — with data-flow diagrams.

CPU and GPU training pipelines, loss functions, gradient flow, and TargetProp (chain-rule and gap-based modes).

Numerical Types

All 21 DTypes from float64 to binary, WeightStore lifecycle, SimulatePrecision, Q4_0, and compression ratios.

DispatchLayer routing, 3D grid traversal, tiled parallel execution, IsRemoteLink spatial hopping, and the GPU dispatch path.

WGPUContext, InitWGPU, BeginFrame/FlushFrame, buffer management, bind group cache, GPU support matrix, WGSL shader overview.

The systolic grid engine: SystolicState, one-clock-cycle forward, spatial feedback via remote links, BPTT, online learning.

Topological network fingerprinting: ExtractDNA, CosineSimilarity, CompareNetworks, LogicShift detection.

Evolution & NEAT

DNA Splice, genetic crossover, and NEAT-style topology evolution: SpliceDNA, NEATMutate, NEATPopulation, all mutation types.

MHA with RoPE, GQA/MQA, KV cache, SwiGLU, RMSNorm, full transformer block assembly, and Transformer[T] generation type.

Full save/load (SerializeNetwork/DeserializeNetwork), bit-packing formats, idempotency guarantee, SafeTensors support.

PTQ pipeline, WeightStore versioning, Morph/Unpack, Q4_0Block block quantization, calibration, accuracy trade-offs.

Softmax Variants

All 10 softmax variants: Standard, Temperature, Gumbel, Masked, Sparse, Entmax, Grid, Hierarchical, Adaptive, Mixture.

Parallel & Sequential

LayerParallel (5 combine modes, activation tree), LayerSequential (step containers, skip gradients), nesting patterns.

Deployment & Polyglot

Polyglot Ecosystem: NPM deployment, TypeScript SDK, WASM bridge, Browser and Node.js usage.