a tiny tensor computation library

What is TinyTensor?

TinyTensor is a lightweight tensor library written from scratch in pure C, focused on building the core foundations of deep learning frameworks. It provides basic tensor structures, memory handling, and essential operations with a minimal and educational design.

DType supported

DType	Description	CPU	CUDA
int8	8bit signed integer	✅	✅
uint8	8bit unsigned integer	✅	✅
int16	16bit signed integer	✅	✅
uint16	16bit unsigned integer	✅	✅
int32	32bit signed integer	✅	✅
uint32	32bit unsigned integer	✅	✅
int64	64bit signed integer	✅	✅
uint64	64bit unsigned integer	✅	✅
float16	16bit floating point	✅	✅
float32	32bit floating point	✅	✅
float64	64bit floating point	✅	✅
float128	128bit floating point	❌	❌
complex64	64bit complex dtype	✅	✅
complex128	128bit complex dtype	✅	✅
complex256	256bit complex dtype	❌	❌

Install from the source

git clone https://github.com/0xhilSa/tinytensor
cd tinytensor
bash build.sh

Testing

import tinytensor as tt

x = tt.Tensor([[1,2,3,4],[5,6,7,8]], dtype=tt.dtypes.int8, device="cpu", const=True)
y = tt.Tensor([[1,2,3,4],[5,6,7,8]], dtype=tt.dtypes.int8, device="cuda", const=True)
print("=====on CPU=====")
print(x)
print(x.buf)
print(x.device)
print("=====on CUDA=====")
print(y)
print(y.buf)
print(y.device)

or simply run pytest -v

Requirements

• GCC / Clang
• CUDA Toolkit (for GPU support)
• Python ≥ 3.9
• Linux environment

Contributions

Pull requests are welcome. If you find a bug or want to suggest an operation, feel free to open an issue.

LICENSE

MIT