tensorax 0.2.0

A high-performance tensor library with CUDA acceleration

⚡ Tensorax

A from-scratch tensor library with hand-written CUDA kernels.

No PyTorch. No NumPy. Pure C++/CUDA + Python.

Usage Guide · Architecture · Contributing · Examples

---

🔩   Zero heavy dependencies Only pybind11 — no PyTorch, NumPy, or cuBLAS at runtime. ⚡   Hand-written CUDA kernels 6 matmul variants, 4 attention kernels, 14 element-wise ops — all from scratch. 🧠   Full autograd engine Reverse-mode autodiff with gradient tracking through 18+ operations.

🎯   PyTorch-like API Familiar Tensor, nn.Module, optim.Adam interface — minimal learning curve. 🧱   Batteries included Linear, ReLU, LayerNorm, BatchNorm, Dropout, GQA, Flash Attention — ready to train. 📚   Built to learn from Clean, readable implementation of a DL framework from first principles.

---

Get Started

pip install tensorax

from tensorax import Tensor, nn, optim, functional as F

# Build
model = nn.Sequential(nn.Linear(4, 8), nn.ReLU(), nn.LayerNorm(8), nn.Linear(8, 3))
optimizer = optim.Adam(model.parameters(), lr=0.001)

# Train
for epoch in range(100):
    loss = F.mse_loss(model(x_train), y_train)
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

→ Full usage guide with all APIs, code examples, and details: docs/USAGE.md

---

What's Inside

Core Tensor with CPU ↔ CUDA Broadcasting arithmetic sum, mean with keepdim reshape, transpose exp, log, sqrt, pow 13 dtype constants	Neural Networks Linear, Sequential ReLU, Sigmoid, Tanh, Softmax LayerNorm, RMSNorm, BatchNorm Dropout Module base class	Training SGD with momentum Adam with bias correction MSE, CrossEntropy, CE from logits Autograd through 18+ ops
Attention Scaled dot-product attention 4 CUDA kernels (naive → flash) Grouped Query Attention Causal & padding masks	CUDA Kernels 6 matmul implementations 14 element-wise ops Parallel reductions Tiled + coalesced access	Infra 400 tests, 98% coverage CI/CD with GitHub Actions pybind11 bindings Automatic CUDA fallback

---

Performance

Matrix Multiplication — fp32, 3×1024×1024, 100 runs:

PyTorch CUDA (ref)         ████████████████████████████████████████████  0.41s  (4.51×)
Tensorax 1D Block Tiling   ██████████████████████████████████████████    0.95s  (2.31×)  ← best
Tensorax Tiled             ████████████████████████████████              1.22s  (1.80×)
NumPy CPU (baseline)       █████████████████████████                    1.85s  (1.00×)

2.31× faster than NumPy · 43% of PyTorch's cuBLAS kernels · all hand-written, zero library calls

Attention Kernels — 4 implementations from naive to flash, supporting arbitrary batch/heads, asymmetric sequence lengths, and optional masks.

---

Project Structure

csrc/                           C++ / CUDA backend
  cuda/kernels/                   elementwise · matmul (×6) · reduction · attention (×4)
  cpu/                            CPU fallback for all ops
  tensor_ops.{cpp,h}             pybind11 bindings

tensorax/                       Python package
  tensor.py                       Tensor class + autograd (1100 lines)
  functional.py                   F.relu, F.softmax, F.sdpa, ...
  nn/                             Linear, norms, dropout, attention, GQA
  optim.py                        SGD, Adam

---

Roadmap

Status
✅	Core ops · autograd · NN layers · norms · optimizers · losses · attention (4 CUDA kernels) · GQA · matmul (6 variants)
🚧	Multi-head attention with projections · expanded benchmarking
🔮	Conv2D · MaxPool2D · GELU/Swish · AdamW · LR schedulers · indexing/slicing · serialization · multi-GPU · mixed precision · DDP

---

Documentation

Usage Guide	API reference, code examples, training patterns
Architecture	System design, kernel strategy, autograd internals
Development	Build, test, contribute
Examples	Runnable scripts for common tasks

---

Contributing

Fork → Branch → Commit → PR

See DEVELOPMENT.md for build instructions and guidelines.

---

Citation

@software{tensorax2025,
  title  = {Tensorax: Pure C++/CUDA Tensor Library},
  author = {Shrirang Mahajan},
  year   = {2025},
  url    = {https://github.com/NotShrirang/tensorax}
}

---

GitHub  ·  Issues  ·  Discussions

Built with ❤️ by @NotShrirang

⭐ Star if you find this useful