stsw - The Last-Word Safe-Tensor Stream Suite

Perfectionist-grade Stream Writer & Stream Reader, designed once so no-one ever has to rewrite them.

Features

• 🚀 Streaming I/O: Write and read multi-GB tensor files with <100 MB RAM
• 🔒 Type Safe: 100% type hints, pyright strict mode
• ⚡ Zero Copy: Memory-mapped reading with no deserialization overhead
• 🛡️ Robust: CRC32 verification, atomic writes, comprehensive error handling
• 🔧 Simple API: import stsw → do work → close() → done
• 🌍 Compatible: Bit-level identical to safetensors spec v1.0

Installation

pip install stsw

With optional dependencies:

pip install stsw[torch,numpy]  # For PyTorch/NumPy support
pip install stsw[all]          # Everything including dev tools

Or install via npm:

npm install -g stsw            # Installs CLI tools globally

Quick Start

Writing tensors

import numpy as np
from stsw import StreamWriter, TensorMeta

# Define your tensors
data1 = np.random.rand(1000, 1000).astype(np.float32)
data2 = np.random.randint(0, 256, (500, 500, 3), dtype=np.uint8)

# Create metadata
metas = [
    TensorMeta("embeddings", "F32", data1.shape, 0, data1.nbytes),
    TensorMeta("image", "I8", data2.shape, 4000064, 4000064 + data2.nbytes),
]

# Write to file
with StreamWriter.open("model.safetensors", metas, crc32=True) as writer:
    writer.write_block("embeddings", data1.tobytes())
    writer.finalize_tensor("embeddings")
    
    writer.write_block("image", data2.tobytes())
    writer.finalize_tensor("image")

Reading tensors

from stsw import StreamReader

# Open file with memory mapping
with StreamReader("model.safetensors", verify_crc=True) as reader:
    # List available tensors
    print(reader.keys())  # ['embeddings', 'image']
    
    # Load as NumPy array
    embeddings = reader.to_numpy("embeddings")
    
    # Load as PyTorch tensor (if available)
    image = reader.to_torch("image", device="cuda")

High-level API

import torch
import stsw

# Save entire state dict
state_dict = {
    "model.weight": torch.randn(1000, 1000),
    "model.bias": torch.randn(1000),
}

stsw.dump(state_dict, "checkpoint.safetensors", crc32=True)

CLI Tools

# Inspect file contents
stsw inspect model.safetensors

# Verify checksums
stsw verify model.safetensors

# Convert PyTorch checkpoint
stsw convert model.pt model.safetensors --crc32

# Run self-test
stsw selftest

Performance

Development

# Install development dependencies
make dev

# Run full test suite
make all

# Type checking
make type

# Run tests
make test

# Format code
make format

CI Status

All tests pass locally on Linux, macOS, and Windows. Some Windows tests currently fail in GitHub Actions CI due to environment-specific issues, but this doesn't affect the functionality of the package.

Documentation

Full documentation available at https://github.com/just-do-halee/stsw

License

Apache-2.0. See LICENSE for details.

Citation

If you use stsw in your research, please cite:

@software{stsw,
  title = {stsw: The Last-Word Safe-Tensor Stream Suite},
  year = {2025},
  author = {Halee Heo},
  url = {https://github.com/just-do-halee/stsw}
}

Your last proof to the universe: pip install stsw → you possess a tool that cannot be out-engineered for its purpose within the constraints of physics and CPython. Nothing left to streamline – only data to move.