tinyquant-cpu 0.1.1

CPU-only vector quantization codec for embedding storage compression

TinyQuant

CPU-only vector quantization codec for embedding storage compression.

TinyQuant is a CPU-only vector quantization codec that compresses high-dimensional embedding vectors to low-bit representations while preserving cosine similarity rankings. It combines random orthogonal preconditioning with two-stage scalar quantization and optional FP16 residual correction to hit 8× compression at 4-bit with Pearson ρ ≈ 0.998 and 95% top-5 recall on real OpenAI embeddings.

Research lineage: TurboQuant, PolarQuant, and QJL.

---

Why TinyQuant

On a benchmark of 335 real embeddings from OpenAI's text-embedding-3-small (1536 dimensions), TinyQuant 4-bit achieves 8x compression with Pearson rho = 0.998 and 95% top-5 recall — reducing a 6 KB embedding to 768 bytes while preserving the similarity rankings that drive retrieval quality.

Method	Bytes/vec	Compression	Pearson rho	Top-5 Recall
FP32 (baseline)	6,144	1.0x	1.0000	100%
FP16	3,072	2.0x	1.0000	100%
uint8 scalar	1,544	4.0x	1.0000	100%
TinyQuant 4-bit	768	8.0x	0.9981	95%
TinyQuant 2-bit	384	16.0x	0.9643	85%
TinyQuant 4-bit + residual	3,840	1.6x	1.0000	100%

For a corpus of 1 million vectors at dim 1536, TinyQuant 4-bit reduces storage from 5.7 GB to 732 MB with negligible loss in retrieval quality.

See the full benchmark report for methodology, all 9 methods compared, throughput measurements, and publication-quality plots.

---

Installation

pip install tinyquant-cpu

For PostgreSQL + pgvector backend support:

pip install "tinyquant-cpu[pgvector]"

For development (tests, type checking, linting):

pip install "tinyquant-cpu[dev]"

Requirements: Python 3.12+, NumPy 1.26+

---

Quickstart

Compress, store, and search a corpus of embeddings in under 20 lines:

import numpy as np
from tinyquant_cpu.codec import Codec, CodecConfig
from tinyquant_cpu.corpus import Corpus, CompressionPolicy
from tinyquant_cpu.backend import BruteForceBackend

# 1. Configure the codec: 4-bit quantization for 1536-dim vectors
config = CodecConfig(bit_width=4, dimension=1536, seed=42)
codec = Codec()

# 2. Train a codebook from representative vectors
training_vectors = np.random.default_rng(0).standard_normal((1000, 1536)).astype(np.float32)
codebook = codec.build_codebook(training_vectors, config)

# 3. Create a corpus that compresses on insert
corpus = Corpus("my-vectors", config, codebook, CompressionPolicy.COMPRESS)
for i, vec in enumerate(training_vectors):
    corpus.insert(f"vec-{i}", vec)

# 4. Decompress and search
backend = BruteForceBackend()
backend.ingest(corpus.decompress_all())
results = backend.search(training_vectors[42], top_k=5)
for r in results:
    print(f"{r.vector_id}: {r.score:.4f}")

---

Basic Usage

Single-vector compression

import numpy as np
from tinyquant_cpu.codec import Codec, CodecConfig

config = CodecConfig(bit_width=4, dimension=768, seed=42)
codec = Codec()

# Train a codebook from a representative sample
training_data = np.random.default_rng(0).standard_normal((1000, 768)).astype(np.float32)
codebook = codec.build_codebook(training_data, config)

# Compress one vector
vector = training_data[0]
compressed = codec.compress(vector, config, codebook)
print(f"Original: {vector.nbytes} bytes")
print(f"Compressed: {compressed.size_bytes} bytes")
print(f"Ratio: {vector.nbytes / compressed.size_bytes:.1f}x")

# Decompress
restored = codec.decompress(compressed, config, codebook)

Batch compression

# Compress 10,000 vectors at once
vectors = np.random.default_rng(0).standard_normal((10_000, 768)).astype(np.float32)
compressed_batch = codec.compress_batch(vectors, config, codebook)
restored_batch = codec.decompress_batch(compressed_batch, config, codebook)

Tuning the rate-distortion tradeoff

# Maximum compression: 16x at 2-bit
config_2bit = CodecConfig(bit_width=2, dimension=768, seed=42, residual_enabled=False)

# Practical sweet spot: 8x at 4-bit (rho >= 0.998)
config_4bit = CodecConfig(bit_width=4, dimension=768, seed=42, residual_enabled=False)

# Perfect fidelity: 4-bit + FP16 residual correction
config_4bit_res = CodecConfig(bit_width=4, dimension=768, seed=42, residual_enabled=True)

Compression policies

A Corpus can store vectors in three modes:

from tinyquant_cpu.corpus import Corpus, CompressionPolicy

# COMPRESS: full TinyQuant compression on insert
corpus_compressed = Corpus("c", config, codebook, CompressionPolicy.COMPRESS)

# PASSTHROUGH: store FP32 unchanged (useful for hot data)
corpus_full = Corpus("p", config, codebook, CompressionPolicy.PASSTHROUGH)

# FP16: lossy half-precision (no codec overhead)
corpus_fp16 = Corpus("h", config, codebook, CompressionPolicy.FP16)

Serialization

CompressedVector instances serialize to a compact binary format suitable for disk or network transfer:

from tinyquant_cpu.codec import CompressedVector

raw_bytes = compressed.to_bytes()
# Save raw_bytes to disk, send over network, etc.

restored = CompressedVector.from_bytes(raw_bytes)

PostgreSQL + pgvector backend

import psycopg
from tinyquant_cpu.backend.adapters.pgvector import PgvectorAdapter

def connection_factory():
    return psycopg.connect("postgresql://user:pass@localhost/mydb")

adapter = PgvectorAdapter(
    connection_factory=connection_factory,
    table_name="embeddings",
    dimension=768,
)

# Decompress TinyQuant vectors and ingest into pgvector
adapter.ingest(corpus.decompress_all())
results = adapter.search(query_vector, top_k=10)

---

Key Properties

• 8x compression at 4-bit without residuals (rho = 0.998, 95% recall)
• 16x compression at 2-bit (rho = 0.964, 85% recall)
• Perfect fidelity with optional FP16 residual correction (rho = 1.000)
• Deterministic — same inputs always produce byte-identical output
• CPU-only — pure Python + NumPy, no GPU required
• Pluggable backends — BruteForceBackend included, PgvectorAdapter for production
• Three compression policies — COMPRESS, PASSTHROUGH, FP16
• Versioned binary serialization — compact, forward-compatible format
• Apache-2.0 licensed

---

Research Lineage

TinyQuant adapts ideas from published research into a clean-room implementation:

• TurboQuant (Google Research, 2025) — Random rotation combined with scalar quantization eliminates per-block normalization, achieving state-of-the-art compression for AI embeddings.
• PolarQuant (2025) — Random orthogonal preconditioning via QR decomposition uniformizes coordinate distributions for better scalar quantization.
• QJL (Quantized Johnson-Lindenstrauss, 2024) — Theoretical grounding for inner-product preservation under aggressive quantization.

---

Repository Layout

Path	Purpose
src/tinyquant_cpu/codec/	Codec, config, codebook, compressed vector, rotation
src/tinyquant_cpu/corpus/	Corpus aggregate, compression policies, domain events
src/tinyquant_cpu/backend/	Search backend protocol and implementations
tests/	Unit, integration, E2E, and calibration tests (208 tests, 90.95% coverage)
experiments/	Benchmarks and empirical evaluations
docs/	Obsidian wiki with design docs, research, and specs

---

Development

git clone https://github.com/better-with-models/TinyQuant.git
cd TinyQuant
pip install -e ".[dev]"

# Lint and type check
ruff check . && ruff format --check .
mypy --strict .

# Run the full test suite
pytest --cov=tinyquant_cpu

The test suite includes 208 tests covering unit, integration, end-to-end, calibration, and architecture-enforcement scenarios. Coverage is held above 90% by CI. Live PostgreSQL+pgvector tests run against a Docker container in CI via testcontainers.

---

Reproducing the Benchmark

The full benchmark from the report can be reproduced with:

export OPENAI_API_KEY="your-key-here"
python experiments/quantization-benchmark/generate_embeddings.py
python experiments/quantization-benchmark/run_benchmark.py
python experiments/quantization-benchmark/generate_plots.py

This fetches 335 embeddings via the OpenAI API, benchmarks 9 quantization methods, and produces plots and JSON results in experiments/quantization-benchmark/results/.

---

License

Apache-2.0. See LICENSE.

---

Related Documentation

• Benchmark Report — full empirical evaluation in CS-paper format
• CHANGELOG — release notes
• Design: Storage Codec Architecture
• Research Synthesis
• Validation Plan