tenso 0.9.0

High-performance zero-copy tensor protocol

Tenso

Up to 23.8x faster than Apache Arrow. 61x less CPU than SafeTensors.

Zero-copy, SIMD-aligned tensor protocol for high-performance ML infrastructure.

---

Why Tenso?

Most serialization formats are designed for general data or disk storage. Tenso is focused on network tensor transmission where every microsecond matters.

The Problem

Traditional formats waste CPU cycles:

• SafeTensors: 36.7% CPU usage per deserialization (great for disk, overkill for network)
• Pickle: 41.7% CPU usage + security vulnerabilities
• Arrow: Fast, but 23.8x slower than Tenso for large tensors

The Solution

Tenso achieves true zero-copy with:

• Fixed 8-byte header (no JSON parsing overhead)
• 64-byte memory alignment (SIMD-ready)
• Direct memory mapping (CPU just points, never copies)

Result: 0.6% CPU usage vs 36.7% for SafeTensors

---

Benchmarks

System: Python 3.12.9, NumPy 2.3.5, 12 CPU cores, macOS

Deserialization Speed (8192×8192 Float32 Matrix)

Format	Time	CPU Usage	Speedup
Tenso	0.034ms	0.6%	1x
Arrow	0.805ms	1.1%	23.8x slower
SafeTensors	2.621ms	36.7%	77x slower
Pickle	3.293ms	41.7%	97x slower

Stream Reading Performance (95MB Packet)

Method	Time	Throughput	Speedup
Tenso read_stream	21ms	4,500 MB/s	1x
Naive loop	7,870ms	12 MB/s	371x slower

Network Latency (1KB Tensor over TCP)

Metric	Value
Throughput	182,940 packets/sec
Latency	5.5 μs/packet

Real-World Impact

Scenario: Inference API serving 10,000 req/sec with 64MB tensors

Format	CPU Cores Used	Monthly Cost*
SafeTensors	367 cores	~$15,000
Tenso	6 cores	~$245

*Based on typical cloud compute pricing

---

Installation

pip install tenso

---

Quick Start

Basic Serialization

import numpy as np
import tenso

# Create tensor
data = np.random.rand(1024, 1024).astype(np.float32)

# Serialize (8.5ms for 64MB)
packet = tenso.dumps(data)

# Deserialize (0.034ms for 64MB)
restored = tenso.loads(packet)

Network Communication

import socket
import tenso

# Server: Receive tensor
server = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
server.bind(('0.0.0.0', 9999))
server.listen(1)
conn, addr = server.accept()

# Zero-copy read with automatic buffering
tensor = tenso.read_stream(conn)  # Uses readinto() internally

# Client: Send tensor
client = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
client.connect(('localhost', 9999))

data = np.random.rand(256, 256).astype(np.float32)
tenso.write_stream(data, client)  # Atomic write with os.writev

File I/O with Memory Mapping

# Write to disk
with open("model_weights.tenso", "wb") as f:
    tenso.dump(large_tensor, f)

# Instant load (no matter the size)
with open("model_weights.tenso", "rb") as f:
    weights = tenso.load(f, mmap_mode=True)  # Memory-mapped, not loaded into RAM

---

Use Cases

Perfect For

• Model Serving APIs - 23.8x faster deserialization saves CPU cores
• Distributed Training - Efficient gradient/activation passing (Ray, Spark)
• Real-time Robotics - Sub-millisecond latency sensor fusion
• High-Frequency Trading - Microsecond-precision data exchange
• Microservices - Fast tensor exchange between services
• Edge Devices - Minimal dependencies, pure Python

Consider Alternatives For

• Long-term Model Storage - Use SafeTensors (better ecosystem, HuggingFace integration)
• Multi-column Dataframes - Use Arrow (designed for tabular data)
• Arbitrary Python Objects - Use Pickle (if you trust the source)

---

Protocol Design

Tenso uses a minimalist 4-part structure:

┌─────────────┬──────────────┬──────────────┬────────────────────────┬──────────────┐
│   HEADER    │    SHAPE     │   PADDING    │    BODY (Raw Data)     │    FOOTER    │
│   8 bytes   │  Variable    │   0-63 bytes │   C-Contiguous Array   │   8 bytes*   │
└─────────────┴──────────────┴──────────────┴────────────────────────┴──────────────┘
                                                                        (*Optional)

Header (8 bytes)

[4 bytes: Magic "TNSO"]
[1 byte:  Protocol Version (2)]
[1 byte:  Flags (Bit 0: Aligned, Bit 1: Integrity)]
[1 byte:  Dtype Code]
[1 byte:  Number of Dimensions]

Why This Is Fast

SafeTensors: Uses JSON header - 3.67ms parsing overhead
Arrow: Complex IPC format with schema validation - 0.805ms overhead
Tenso: Fixed 8-byte struct - 0.034ms (just unpack and memory map)

The padding ensures the data body starts at a 64-byte boundary, enabling:

• AVX-512 vectorization
• Zero-copy memory mapping
• Cache-line alignment

---

Advanced Features

Data Integrity (XXH3)

Protect your tensors against network corruption with ultra-fast XXH3 hashing (adds <2% overhead):

# Serialize with 64-bit checksum
packet = tenso.dumps(data, check_integrity=True)

# Verification is automatic during load
try:
    restored = tenso.loads(packet)
except ValueError:
    print("Detected corrupted data!")

Strict Mode

Prevents accidental memory copies:

# Force C-contiguous check
try:
    packet = tenso.dumps(fortran_array, strict=True)
except ValueError:
    print("Array must be C-contiguous!")
    fortran_array = np.ascontiguousarray(fortran_array)

Packet Introspection

Inspect metadata without deserializing:

info = tenso.get_packet_info(packet)
print(f"Shape: {info['shape']}")
print(f"Dtype: {info['dtype']}")
print(f"Size: {info['data_size_bytes']} bytes")

Supported Dtypes

All NumPy numeric types including:

• Floats: float16, float32, float64
• Integers: int8, int16, int32, int64, uint8, uint16, uint32, uint64
• Complex: complex64, complex128
• Boolean: bool

---

Comparison Table

Feature	Tenso	Arrow	SafeTensors	Pickle
Deserialize Speed (64MB)	0.034ms	0.805ms	2.621ms	3.293ms
CPU Usage	0.6%	1.1%	36.7%	41.7%
Memory Overhead	0.00%	0.00%	0.00%	0.00%
Security	Safe	Safe	Safe	RCE Risk
Dependencies	NumPy only	PyArrow (large)	Rust bindings	Python stdlib
Best For	Network/IPC	Dataframes	Disk storage	Python objects
SIMD Aligned	64-byte	64-byte	No	No

---

Performance Deep-Dive

Read the full story: Breaking the Speed Limit: Optimizing Python Tensor Serialization to 5 GB/s

Key insights:

• Why JSON headers kill performance
• How memory alignment enables zero-copy
• Why Tenso beats Arrow for single tensors
• Real-world cost savings ($15k/month at scale)

---

Development

# Clone repository
git clone https://github.com/Khushiyant/tenso.git
cd tenso

# Install with dev dependencies
pip install -e ".[dev]"

# Install with gpu dependencies
pip install -e ".[gpu]"

# Run tests
pytest

# Run comprehensive benchmarks
python benchmark.py all

# Quick benchmark (serialization + Arrow comparison)
python benchmark.py quick

# Benchmark with Integrity
python benchmark.py all --integrity # Require installation with [integrity] dependicies

---

Requirements

• Python >= 3.10
• NumPy >= 1.20

Optional (for benchmarks):

• pyarrow - Compare with Apache Arrow

• safetensors - Compare with SafeTensors

• msgpack - Compare with MessagePack

• psutil - Monitor CPU/memory usage

• xxhash - Integrity Checks Implementation

---

Contributing

Contributions welcome. Areas we'd love help with:

• Async support (async def aread_stream())
• Compression integration (zstd, lz4)
• gRPC/FastAPI integration examples
• Rust bindings for even faster serialization
• JavaScript/WASM client for browser ML
• CUDA support for GPU-direct transfers

---

License

MIT License - see License file.

---

Citation

If you use Tenso in research, please cite:

@software{tenso2025,
  author = {Khushiyant},
  title = {Tenso: High-Performance Zero-Copy Tensor Protocol},
  year = {2025},
  url = {https://github.com/Khushiyant/tenso}
}

---

Acknowledgments

Inspired by the need for faster ML inference infrastructure. Built with care for the ML community.

Star this repo if Tenso saved you CPU cycles.