https://github.com/invrs-io/jeig
Faster eigendecompositions for jax
https://github.com/invrs-io/jeig
Last synced: 7 months ago
JSON representation
Faster eigendecompositions for jax
- Host: GitHub
- URL: https://github.com/invrs-io/jeig
- Owner: mfschubert
- License: bsd-3-clause
- Created: 2024-06-11T15:47:40.000Z (about 2 years ago)
- Default Branch: main
- Last Pushed: 2025-08-26T17:48:01.000Z (11 months ago)
- Last Synced: 2025-08-26T23:14:08.352Z (11 months ago)
- Language: Python
- Homepage:
- Size: 164 KB
- Stars: 8
- Watchers: 1
- Forks: 1
- Open Issues: 2
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# jeig - Eigendecompositions wrapped for jax
[](https://github.com/invrs-io/jeig/actions)
[](https://pypi.org/project/jeig/)
## Overview
This package wraps eigendecompositions as provided by jax, cusolver, magma, numpy, scipy, and torch for use with jax. Depending upon your system and your versions of these packages, you may observe significant speed differences. The following were obtained using jax 0.8.0 on a system with 28-core Intel Xeon w7-3465X and NVIDIA RTX4090.
## Install
jeig can be installed via pip,
```
pip install jeig
```
This will also install torch. If you only need torch for use with jeig, then the CPU-only version could be sufficient and you may wish to install manually as described in the [pytorch docs](https://pytorch.org/get-started/locally/).
## Example usage
```python
import jax
import jeig
matrix = jax.random.normal(jax.random.PRNGKey(0), (1, 2048, 2048)).astype(complex)
%timeit jax.block_until_ready(jeig.eig(matrix, backend="cusolver"))
%timeit jax.block_until_ready(jeig.eig(matrix, backend="lapack"))
%timeit jax.block_until_ready(jeig.eig(matrix, backend="magma"))
%timeit jax.block_until_ready(jeig.eig(matrix, backend="torch"))
```
```
1.31 s ± 43 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
5.44 s ± 379 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
11.1 s ± 937 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
4.93 s ± 92.1 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
```
The default torch backend has good performance when performing batched eigendecomposition on many-core CPUs.
```python
matrix = jax.random.normal(jax.random.PRNGKey(0), (8, 2048, 2048)).astype(complex)
%timeit jax.block_until_ready(jeig.eig(matrix, backend="cusolver"))
%timeit jax.block_until_ready(jeig.eig(matrix, backend="lapack"))
%timeit jax.block_until_ready(jeig.eig(matrix, backend="magma"))
%timeit jax.block_until_ready(jeig.eig(matrix, backend="torch"))
```
```
10.4 s ± 116 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
48.1 s ± 6.74 s per loop (mean ± std. dev. of 7 runs, 1 loop each)
1min 33s ± 1.49 s per loop (mean ± std. dev. of 7 runs, 1 loop each)
7.18 s ± 91.6 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
```
## Credit
The torch implementation of eigendecomposition is due to a [comment](https://github.com/google/jax/issues/10180#issuecomment-1092098074) by @YouJiacheng.