ztensor 1.2.3

Unified, zero-copy, and safe I/O for deep learning formats
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62

# ztensor

[![PyPI](https://img.shields.io/pypi/v/ztensor.svg)](https://pypi.org/project/ztensor/)
[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)

Python bindings for the [zTensor](https://github.com/pie-project/ztensor) tensor serialization library.

## Installation

```bash
pip install ztensor
```

**Optional dependencies:**
- `torch` — PyTorch support
- `scipy` — Reading sparse tensors as scipy sparse matrices
- `ml_dtypes` — NumPy bfloat16 support (`pip install ztensor[bfloat16]`)

## Quick Start

```python
import numpy as np
from ztensor import Writer, Reader

# Write tensors
with Writer("model.zt") as w:
    w.add("weights", np.random.randn(1024, 768).astype(np.float32))
    w.add("bias", np.zeros(768, dtype=np.float32))

# Read tensors (zero-copy mmap)
with Reader("model.zt") as r:
    weights = r["weights"]  # numpy array
    print(f"Shape: {weights.shape}, dtype: {weights.dtype}")
```

zTensor can also read `.safetensors`, `.pt`, `.gguf`, `.npz`, `.onnx`, and `.h5` files through the same API — format detection is automatic:

```python
with Reader("model.safetensors") as r:
    weights = r["weights"]
```

## PyTorch

```python
from ztensor.torch import save_file, load_file

save_file({"embed": torch.randn(1000, 768)}, "model.zt")
tensors = load_file("model.zt", device="cuda:0")
```

## Supported Data Types

`float64`, `float32`, `float16`, `bfloat16`, `int64`, `int32`, `int16`, `int8`, `uint64`, `uint32`, `uint16`, `uint8`, `bool`

## Documentation

See the [full API reference](https://pie-project.github.io/ztensor/python) for details on Writer, Reader, sparse tensors, compression, and more.

## License

MIT