# Supersayan
A high‑performance Python library that integrates Fully Homomorphic Encryption (FHE) into PyTorch workflows. Supersayan provides:
- A client/server path to offload selected layers to a remote FHE server (hybrid mode).
- Zero‑copy tensor bridges across PyTorch, NumPy, CuPy, and Julia.
The Julia backend is managed via `juliacall` and initialized automatically on first import, or explicitly with a helper CLI.
## Installation
From PyPI:
```bash
pip install supersayan
# or with uv
uv add supersayan
```
Default behavior:
- After install, importing `supersayan` triggers a one‑time Julia backend setup automatically. No extra step is needed for typical users.
Optional (CI/Docker or troubleshooting):
```bash
# Manually initialize the Julia backend if you want explicit control
supersayan-setup
```
Advanced control (not recommended for regular users):
- To prevent network access during import in CI images, set `SUPERSAYAN_SKIP_JULIA_SETUP=1` and run `supersayan-setup` in a controlled build step.
Notes:
- GPU support uses CuPy when available; code runs on CPU otherwise.
## Hybrid Remote Inference
Run the TCP server:
```bash
python scripts/run_server.py --host 127.0.0.1 --port 8000 --models-dir /tmp/supersayan/models
```
Use the client to execute only selected layers remotely in FHE while keeping other ops local:
```python
import torch
import torch.nn as nn
from supersayan.core.types import SupersayanTensor
from supersayan.remote import SupersayanClient
class SmallCNN(nn.Module):
def __init__(self):
super().__init__()
self.conv = nn.Conv2d(1, 4, 3, padding=1)
self.fc = nn.Linear(4*28*28, 10)
def forward(self, x):
x = torch.relu(self.conv(x))
return self.fc(x.view(x.size(0), -1))
cnn = SmallCNN().eval()
client = SupersayanClient(
server_url="127.0.0.1:8000",
torch_model=cnn,
fhe_modules=[nn.Conv2d, nn.Linear], # offload these layers
)
x = SupersayanTensor(torch.randn(1, 1, 28, 28))
y = client(x)
```
For a runnable example of the TCP server, see `scripts/run_server.py`.
Supported offloaded layers: `nn.Linear`, `nn.Conv2d`.
## Tensors and Interop
- `SupersayanTensor(data, device=...)` accepts `torch.Tensor`, `numpy.ndarray`, or `cupy.ndarray` and preserves dtype float32.
- Helpers: `SupersayanTensor.zeros(...)`, `ones(...)`, `randn(...)`.
- Interop: `.to_numpy()`, `.to_dlpack()`, and zero‑copy conversion to Julia via `.to_julia()`.
## Project Layout
- `src/supersayan/` core, layers, remote client/server, Julia backend
- `scripts/` runnable examples (`run_server.py`)
## Troubleshooting
- If Julia setup fails on first import, run `supersayan-setup` manually.
- In CI or headless environments, set `SUPERSAYAN_SKIP_JULIA_SETUP=1` during import and run `supersayan-setup` explicitly in a build step.
Raw data
{
"_id": null,
"home_page": null,
"name": "supersayan",
"maintainer": null,
"docs_url": null,
"requires_python": ">=3.9",
"maintainer_email": null,
"keywords": "homomorphic-encryption, FHE, deep-learning, pytorch, privacy-preserving, cryptography, neural-networks, secure-computation, machine-learning, julia",
"author": null,
"author_email": "bonsainoodle <3sprix.exe@gmail.com>, franklintra <franklin.tranie@gmail.com>",
"download_url": "https://files.pythonhosted.org/packages/d0/dc/b9e649cfc0406319b311485ac48ea5abf72bd8eaedb8bd3805abc5af7281/supersayan-0.1.3.tar.gz",
"platform": null,
"description": "# Supersayan\n\nA high\u2011performance Python library that integrates Fully Homomorphic Encryption (FHE) into PyTorch workflows. Supersayan provides:\n\n- A client/server path to offload selected layers to a remote FHE server (hybrid mode).\n- Zero\u2011copy tensor bridges across PyTorch, NumPy, CuPy, and Julia.\n\nThe Julia backend is managed via `juliacall` and initialized automatically on first import, or explicitly with a helper CLI.\n\n## Installation\n\nFrom PyPI:\n\n```bash\npip install supersayan\n# or with uv\nuv add supersayan\n```\n\nDefault behavior:\n- After install, importing `supersayan` triggers a one\u2011time Julia backend setup automatically. No extra step is needed for typical users.\n\nOptional (CI/Docker or troubleshooting):\n\n```bash\n# Manually initialize the Julia backend if you want explicit control\nsupersayan-setup\n```\n\nAdvanced control (not recommended for regular users):\n- To prevent network access during import in CI images, set `SUPERSAYAN_SKIP_JULIA_SETUP=1` and run `supersayan-setup` in a controlled build step.\n\nNotes:\n- GPU support uses CuPy when available; code runs on CPU otherwise.\n\n## Hybrid Remote Inference\n\nRun the TCP server:\n\n```bash\npython scripts/run_server.py --host 127.0.0.1 --port 8000 --models-dir /tmp/supersayan/models\n```\n\nUse the client to execute only selected layers remotely in FHE while keeping other ops local:\n\n```python\nimport torch\nimport torch.nn as nn\nfrom supersayan.core.types import SupersayanTensor\nfrom supersayan.remote import SupersayanClient\n\nclass SmallCNN(nn.Module):\n def __init__(self):\n super().__init__()\n self.conv = nn.Conv2d(1, 4, 3, padding=1)\n self.fc = nn.Linear(4*28*28, 10)\n def forward(self, x):\n x = torch.relu(self.conv(x))\n return self.fc(x.view(x.size(0), -1))\n\ncnn = SmallCNN().eval()\nclient = SupersayanClient(\n server_url=\"127.0.0.1:8000\",\n torch_model=cnn,\n fhe_modules=[nn.Conv2d, nn.Linear], # offload these layers\n)\n\nx = SupersayanTensor(torch.randn(1, 1, 28, 28))\ny = client(x)\n```\n\nFor a runnable example of the TCP server, see `scripts/run_server.py`.\n\nSupported offloaded layers: `nn.Linear`, `nn.Conv2d`.\n\n## Tensors and Interop\n\n- `SupersayanTensor(data, device=...)` accepts `torch.Tensor`, `numpy.ndarray`, or `cupy.ndarray` and preserves dtype float32.\n- Helpers: `SupersayanTensor.zeros(...)`, `ones(...)`, `randn(...)`.\n- Interop: `.to_numpy()`, `.to_dlpack()`, and zero\u2011copy conversion to Julia via `.to_julia()`.\n\n## Project Layout\n\n- `src/supersayan/` core, layers, remote client/server, Julia backend\n- `scripts/` runnable examples (`run_server.py`)\n\n## Troubleshooting\n\n- If Julia setup fails on first import, run `supersayan-setup` manually.\n- In CI or headless environments, set `SUPERSAYAN_SKIP_JULIA_SETUP=1` during import and run `supersayan-setup` explicitly in a build step.\n",
"bugtrack_url": null,
"license": null,
"summary": "A high-performance Python library for fully homomorphic encryption (FHE) in deep learning, seamlessly integrated with PyTorch.",
"version": "0.1.3",
"project_urls": {
"GitHub": "https://github.com/supersayan-labs/supersayan"
},
"split_keywords": [
"homomorphic-encryption",
" fhe",
" deep-learning",
" pytorch",
" privacy-preserving",
" cryptography",
" neural-networks",
" secure-computation",
" machine-learning",
" julia"
],
"urls": [
{
"comment_text": null,
"digests": {
"blake2b_256": "07ab5f2b8bdc9969ad894fc3c8e9dde60b224d311ceeab6b3d05773ac7db23d4",
"md5": "e1bb56dce297a5c3eadebfe656e167d6",
"sha256": "e803cb8cfe8ceb99d931e64490eeafcc1c054d34e49a85771d814ce94b48f7bc"
},
"downloads": -1,
"filename": "supersayan-0.1.3-py3-none-any.whl",
"has_sig": false,
"md5_digest": "e1bb56dce297a5c3eadebfe656e167d6",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": ">=3.9",
"size": 26374,
"upload_time": "2025-08-28T12:33:05",
"upload_time_iso_8601": "2025-08-28T12:33:05.948713Z",
"url": "https://files.pythonhosted.org/packages/07/ab/5f2b8bdc9969ad894fc3c8e9dde60b224d311ceeab6b3d05773ac7db23d4/supersayan-0.1.3-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": null,
"digests": {
"blake2b_256": "d0dcb9e649cfc0406319b311485ac48ea5abf72bd8eaedb8bd3805abc5af7281",
"md5": "5fbdec954fe48ecb221a34db716c1ed1",
"sha256": "03dd80e34963906b7a66af6b8edbcb115ebfdd6d1e086fad77e5a58bdf1882cb"
},
"downloads": -1,
"filename": "supersayan-0.1.3.tar.gz",
"has_sig": false,
"md5_digest": "5fbdec954fe48ecb221a34db716c1ed1",
"packagetype": "sdist",
"python_version": "source",
"requires_python": ">=3.9",
"size": 22942,
"upload_time": "2025-08-28T12:33:07",
"upload_time_iso_8601": "2025-08-28T12:33:07.121100Z",
"url": "https://files.pythonhosted.org/packages/d0/dc/b9e649cfc0406319b311485ac48ea5abf72bd8eaedb8bd3805abc5af7281/supersayan-0.1.3.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2025-08-28 12:33:07",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "supersayan-labs",
"github_project": "supersayan",
"travis_ci": false,
"coveralls": false,
"github_actions": false,
"lcname": "supersayan"
}
JSON
