Rutorch

Small Rust autograd/tensor playground with CPU + Metal backends, plus learning demos (XOR, decision boundary, CharGPT).

Features

• Autograd on basic tensor ops (CPU + Metal)
• Stable log_softmax + nll_loss
• Toy datasets (spiral/blob) with CSV outputs
• CharGPT mini pipeline (RNN/GRU or minimal GPT)

Quick Start

cargo run --release -- xor

Demos

XOR

cargo run --release -- xor

Spiral / Blob Decision Boundary

cargo run --release -- spiral
cargo run --release -- blob

Outputs:

• out/spiral_points.csv
• out/spiral_grid.csv
• out/blob_points.csv
• out/blob_grid.csv

Python plotting (optional):

python py/spiral_draw.py
python py/blob_draw.py

CharGPT (RNN/GRU)

Train on data/exp.txt (math expressions):

cargo run --release -- char cpu rnn 80
cargo run --release -- char cpu gru 80

GPU on macOS:

cargo run --release -- char gpu rnn 80
cargo run --release -- char gpu gru 80

Arguments:

• device: cpu | gpu
• model: rnn | gru | gpt
• epochs: optional integer (default 80)

CharGPT (GPT)

Minimal single-head causal attention:

cargo run --release -- char cpu gpt 20

Data

data/exp.txt contains 1000 recursive math expressions.
You can regenerate it if needed.

Tests

cargo test -q

Notes

• GPU speedups only help when compute is large and CPU↔GPU transfers are minimized.
• Current RNN/GRU training uses contiguous batch slicing and gradient clipping.