turbo-quant-lite 0.1.0

Numpy-only TurboQuant vector quantization. No PyTorch, no CUDA.

turbo-quant-lite

Numpy-only vector quantization based on Google's TurboQuant algorithm. Compresses float32 vectors to 1-4 bit indices with near-optimal quality. No PyTorch, no CUDA, no model dependencies.

from turbo_quant_lite import TurboQuant

tq = TurboQuant(dim=768, bits=4)

indices, norm = tq.encode(embedding)   # 3072 bytes → 388 bytes
restored = tq.decode(indices, norm)    # < 1.1% MSE distortion

Why this exists

There are two existing TurboQuant implementations on PyPI:

• turboquant — PyTorch-based, focused on LLM KV cache compression. Full HuggingFace integration and GPU support. Requires PyTorch (~2 GB install).
• turboquant-vectors — Numpy-only, focused on batch vector compression and embedding privacy. Includes a PrivateEncoder for protecting embeddings against inversion attacks. Designed for the "compress a collection, save to disk, search the collection" workflow.

This package fills a different niche: per-vector compression for database storage. It's for applications that store embeddings row-by-row in PostgreSQL, SQLite, or Redis and need to compress each vector into a compact binary blob (388 bytes at 4-bit, dim=768) that can be stored in a bytea column or cache key.

Key differences from turboquant-vectors:

• Per-vector binary serialization — pack() / unpack() produce compact bytes for database row storage. No file I/O required.
• Zero per-vector overhead — the quantizer is shared (initialized once), compressed data is just indices + norm. No 2MB wrapper per vector.
• Direct single-vector similarity — similarity(query, indices, norm) works on raw indices without wrapping in a collection object.

Same algorithm, same quality, designed for the database storage use case.

For embedding privacy (protecting against inversion attacks like Vec2Text), see turboquant-vectors PrivateEncoder. You can apply a secret rotation before compression — the two compose naturally as separate layers.

Install

pip install turbo-quant-lite

Or just copy turbo_quant_lite/core.py into your project. It's one file.

What is TurboQuant?

TurboQuant is a data-oblivious vector quantization algorithm from Google Research. It compresses vectors without needing training data or calibration — it works instantly on any vector from any source.

The key insight: randomly rotate a vector and each coordinate becomes approximately Gaussian with known variance. Since the distribution is known in advance, you can precompute the optimal quantization grid. This turns a hard problem (data-dependent codebook learning) into a table lookup.

Results:

• 4-bit: 8x compression, < 1.1% MSE distortion
• 3-bit: ~10x compression, < 4.3% MSE distortion
• 2-bit: 16x compression, < 17% MSE distortion

Quality is within 2.72x of the information-theoretic optimum (Shannon lower bound) at every bit width. This bound is provable and data-independent — it holds for any vector, not just your benchmark.

Paper: Zandieh, Daliri, Hadian, Mirrokni. TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate. ICLR 2026. arXiv:2504.19874

Reference implementations:

• turboquant — PyTorch, KV cache focus, GPU support
• turboquant-rs — Rust, research/verification focus

Usage

Basic encode/decode

import numpy as np
from turbo_quant_lite import TurboQuant

tq = TurboQuant(dim=768, bits=4, seed=42)

# Any float array — from OpenAI, Nebius, Cohere, local model, etc.
embedding = np.random.randn(768).astype(np.float32)

# Compress
indices, norm = tq.encode(embedding)
# indices: uint8 array (768 values, each 0-15 for 4-bit)
# norm: float (the vector's L2 norm)

# Decompress
restored = tq.decode(indices, norm)

# Quality check
mse = np.mean((embedding - restored) ** 2) / np.mean(embedding ** 2)
# mse < 0.011 for 4-bit (guaranteed by theory)

Batch operations

embeddings = np.random.randn(1000, 768)

all_indices, all_norms = tq.encode_batch(embeddings)
restored_batch = tq.decode_batch(all_indices, all_norms)

Approximate similarity (fast, no full decompression)

query = np.random.randn(768)
score = tq.similarity(query, indices, norm)
# Equivalent to cosine similarity but skips the inverse rotation

Binary serialization for storage

from turbo_quant_lite import pack, unpack

# Pack to bytes (388 bytes for 4-bit, dim=768)
data = pack(indices, norm, bits=4)

# Unpack from bytes
indices, norm = unpack(data, dim=768, bits=4)

Use with any embedding provider

# OpenAI
response = openai.embeddings.create(model="text-embedding-3-small", input="hello")
embedding = np.array(response.data[0].embedding)
compressed = tq.encode(embedding)

# Nebius
embedding = await nebius_embedder.embed_text("hello")
compressed = tq.encode(np.array(embedding))

# Sentence Transformers
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("all-MiniLM-L6-v2")
embedding = model.encode("hello")
compressed = tq.encode(embedding)

When to use this

Good fit:

• Storing embeddings in a database (8x size reduction)
• Caching embeddings in Redis/Valkey (8x memory reduction)
• Shipping embeddings over the network (8x bandwidth reduction)
• Local vector search where you control the storage format
• Cold storage / backups of embedding collections
• Edge devices with limited memory

Not a good fit:

• pgvector search (pgvector needs float32/halfvec, no native 4-bit support yet)
• LLM KV cache compression (use turboquant with PyTorch)
• Sub-millisecond latency requirements at dim > 2048 (the rotation matmul becomes the bottleneck)

Performance

On a typical CPU (M-series Mac, modern x86), dim=768:

Operation	Time	Notes
encode (single)	~1.5ms	Dominated by rotation matmul
decode (single)	~1.5ms	Same matmul
encode_batch(1000)	~400ms	Amortized 0.4ms/vector
similarity	~0.3ms	Skips inverse rotation
pack	~0.1ms	Bit packing

Storage sizes (dim=768)

Format	Bytes per vector	Compression
float32	3,072	1x
float16	1,536	2x
4-bit TurboQuant	388	7.9x
3-bit TurboQuant	292	10.5x
2-bit TurboQuant	196	15.7x

Important: seed must match

The rotation matrix is generated from the seed. Encoding with seed=42 and decoding with seed=43 produces garbage. Use the same seed everywhere, or serialize the TurboQuant instance.

License

MIT