Getting Started¶

This guide walks through installation, backend setup, and your first steps with tinyTT.

Installation¶

Prerequisites¶

Python 3.11+
PyTorch 2.0+ — the only backend since 0.5.

Install from Source¶

git clone https://github.com/meigel/tinyTT.git
cd tinyTT
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt
pip install -e .

GPU Support¶

PyTorch handles the device. Pass device= to any constructor, or set TINYTT_DEVICE to run the device-parametrised tests there:

TINYTT_DEVICE=cuda PYTHONPATH=. pytest tests/test_gpu_ops.py -q

Note that float64 is not available on every accelerator; the backend probes with tn.supports_fp64(device) and the tests fall back to float32.

Optional: FEM Problems¶

The problems/ and uq_adf paths need scikit-fem:

pip install -e ".[fem]"

Development Dependencies¶

pip install -r requirements-dev.txt    # pytest, ruff, scipy, coverage, scikit-fem

Your First TT Tensor¶

import numpy as np
import tinytt as tt
import tinytt.backend as tn

# Build a TT tensor from a full 2×2×2 array
full = np.arange(8, dtype=np.float64).reshape(2, 2, 2)
x = tt.TT(full, eps=1e-12)
print("TT ranks:", x.R)                    # [1, 1]
print("Shape:", x.N)                       # [2, 2, 2]

# Materialise back to dense
recon = tn.to_numpy(x.full())
rel_err = np.linalg.norm(recon - full) / np.linalg.norm(full)
print(f"Reconstruction error: {rel_err:.3e}")   # ≈ 0

TT-Matrix and Matvec¶

# Identity TT-matrix
A = tt.eye([4, 4])
x = tt.ones([4, 4])
b = A @ x
print((b - x).norm().numpy())    # ≈ 0

Switching Backends¶

tinyTT has one backend. TINYTT_BACKEND is still read, but only to give a clear error: any value other than pytorch raises ValueError, and the tinygrad case says so by name because 0.4 scripts set it. 0.5 warned and loaded PyTorch anyway; 0.6 refuses, because silently substituting a backend for the one that was asked for is the bug the warning announced. Unset it.

Everything in the package goes through the facade:

import tinytt.backend as tn     # never `import torch`

The facade is where device, dtype and complex-promotion policy live.

Next Steps¶

Follow the TT Basics tutorial for a deeper dive into construction, rounding, and decomposition.
Browse Examples for ready-to-run scripts.
Check the source code on GitHub for detailed module documentation.