Glitch is the primary interactive CLI companion powered by the M-POLY-VTD (Multi-numerical POLYmorphic Volumetric Tiled-tensor Dispatcher) engine. It is a completely pure Go-based neural network architecture executing cross-platform WebGPU computation without any external Native C++ or CUDA dependencies.

Glitch is intentionally designed to be a small, happy robot companion that thinks insults are compliments!

When you start Glitch via go run glitch.go, you are presented with four core tools:

• [1] HuggingFace LLM Mode: Full hardware induction using FlashPoly tiling for local interactive inference. Load any Safetensors model from your HuggingFace cache into instantaneous VRAM.
• [2] Diagnostics & MNIST Simulator: Experimental training and simulator environments for rapid prototyping.
• [3] Testing (Layer Tests): Direct verification of GPU integration (WGPU Buffer Allocation, Dimensional Shapes) across CNN and Dense layers.
• [4] Automated SmolLM2 Test: The Exhaustive Type/Mode Matrix benchmark. This autonomously executes an entire lifecycle of 15 combinations of Numerical Precision (FP32, FP16, BF16, INT8, INT4) and Tiled Execution modes (Standard, Single-Core, Multi-Core) on your active GPU.

Unlike traditional C++ architectures (like llama.cpp) which require specialized CUDA/Metal device bindings, Poly uses dynamic execution of mathematically identical WebGPU Shaders (WGSL), guaranteeing cross-platform consistency.

The engine actively supports "Numerical Metamorphosis," transforming its memory structures per-layer asynchronously, yielding incredible memory footprint shrinks. Below is the real-world performance matrix gathered for SmolLM2-135M-Instruct.

(Note: Data shown is for the INT4 Multi-Core Tiled Forward experimental pass. VRAM heavily optimizes using runtime Tied-Weigh verification mapping).

Glitch includes several experimental demos located in the demos/ directory:

• Audio (conv1d): 1D convolution processing simulations.
• MNIST (conv2d): Small-CNN image classification simulator.
• Time-Series (lstm): LSTM-based sequence prediction.
• SwiGLU: Functional verification of advanced activation layers used in modern LLMs.

Because the engine requires zero C-bindings, it is completely buildable from scratch across platforms. To build Glitch for Android architectures, use the provided assembly script:

# Build for ARM64 (default)
chmod +x android.sh
./android.sh

# Build for x86_64
ARCH=x86_64 ./android.sh

Run the build server script in the glitch directory to serve your compiled/ folder. It will automatically detect your local IP and provide the exact curl command for your Android device.

python3 serve_builds.py