ahb-sjsu/turboquant-pro

Direct quantized cosine similarity (skip decompression) about 3 hours ago

enhancement help wanted good first issue

22

Embedding & KV-cache compression for LLMs and vector databases. PCA-Matryoshka + TurboQuant: 27x compression at 99.8% recall@10 (identical reranking). Larger-than-RAM sharded/memmap search, tqp CLI with distribution-free rank certificates, asymmetric K/V, CUDA+Triton kernels, auto-config. 919 tests. MIT.

Python

Quantization-aware micro-LoRA: rank-8 adapter to recover weight compression quality about 3 hours ago

enhancement help wanted

ahb-sjsu/turboquant-pro

22

Embedding & KV-cache compression for LLMs and vector databases. PCA-Matryoshka + TurboQuant: 27x compression at 99.8% recall@10 (identical reranking). Larger-than-RAM sharded/memmap search, tqp CLI with distribution-free rank certificates, asymmetric K/V, CUDA+Triton kernels, auto-config. 919 tests. MIT.

Python

cuda_search pack_binary + pgvector bytea tests fail on GPU-present, cupy-absent hosts about 3 hours ago

bug good first issue

ahb-sjsu/turboquant-pro

22

Embedding & KV-cache compression for LLMs and vector databases. PCA-Matryoshka + TurboQuant: 27x compression at 99.8% recall@10 (identical reranking). Larger-than-RAM sharded/memmap search, tqp CLI with distribution-free rank certificates, asymmetric K/V, CUDA+Triton kernels, auto-config. 919 tests. MIT.

Python

[Tracking] llama.cpp KV cache integration (5-sprint roadmap) about 3 hours ago

enhancement help wanted

ahb-sjsu/turboquant-pro

22

Embedding & KV-cache compression for LLMs and vector databases. PCA-Matryoshka + TurboQuant: 27x compression at 99.8% recall@10 (identical reranking). Larger-than-RAM sharded/memmap search, tqp CLI with distribution-free rank certificates, asymmetric K/V, CUDA+Triton kernels, auto-config. 919 tests. MIT.

Python

tqvector format does not record the rotation seed (decompress hardcodes 42) about 3 hours ago

bug help wanted

ahb-sjsu/turboquant-pro

22

Embedding & KV-cache compression for LLMs and vector databases. PCA-Matryoshka + TurboQuant: 27x compression at 99.8% recall@10 (identical reranking). Larger-than-RAM sharded/memmap search, tqp CLI with distribution-free rank certificates, asymmetric K/V, CUDA+Triton kernels, auto-config. 919 tests. MIT.

Python

Optimized 4-bit quantizer using binary decision tree about 3 hours ago

enhancement good first issue

ahb-sjsu/turboquant-pro

22

Embedding & KV-cache compression for LLMs and vector databases. PCA-Matryoshka + TurboQuant: 27x compression at 99.8% recall@10 (identical reranking). Larger-than-RAM sharded/memmap search, tqp CLI with distribution-free rank certificates, asymmetric K/V, CUDA+Triton kernels, auto-config. 919 tests. MIT.

Python

Standalone Rust reader for the TQE1 vector and TQIX index formats about 3 hours ago

enhancement help wanted

ahb-sjsu/turboquant-pro

22

Embedding & KV-cache compression for LLMs and vector databases. PCA-Matryoshka + TurboQuant: 27x compression at 99.8% recall@10 (identical reranking). Larger-than-RAM sharded/memmap search, tqp CLI with distribution-free rank certificates, asymmetric K/V, CUDA+Triton kernels, auto-config. 919 tests. MIT.

Python

Freeze the tqvector binary format with golden fixtures + property-based roundtrips about 3 hours ago

enhancement help wanted good first issue

ahb-sjsu/turboquant-pro

22

Embedding & KV-cache compression for LLMs and vector databases. PCA-Matryoshka + TurboQuant: 27x compression at 99.8% recall@10 (identical reranking). Larger-than-RAM sharded/memmap search, tqp CLI with distribution-free rank certificates, asymmetric K/V, CUDA+Triton kernels, auto-config. 919 tests. MIT.

Python

CI: add a cargo/pgrx lane for the pgext PostgreSQL extension about 3 hours ago

enhancement help wanted good first issue

ahb-sjsu/turboquant-pro

22

Embedding & KV-cache compression for LLMs and vector databases. PCA-Matryoshka + TurboQuant: 27x compression at 99.8% recall@10 (identical reranking). Larger-than-RAM sharded/memmap search, tqp CLI with distribution-free rank certificates, asymmetric K/V, CUDA+Triton kernels, auto-config. 919 tests. MIT.

Python

Native FP8 (Hopper) and NVFP4 (Blackwell) kernels 4 months ago

enhancement help wanted

ahb-sjsu/turboquant-pro

22

Embedding & KV-cache compression for LLMs and vector databases. PCA-Matryoshka + TurboQuant: 27x compression at 99.8% recall@10 (identical reranking). Larger-than-RAM sharded/memmap search, tqp CLI with distribution-free rank certificates, asymmetric K/V, CUDA+Triton kernels, auto-config. 919 tests. MIT.

Python

Attention-aware KV eviction: drop or aggressively compress rarely-attended tokens 4 months ago

enhancement help wanted

ahb-sjsu/turboquant-pro

22

Embedding & KV-cache compression for LLMs and vector databases. PCA-Matryoshka + TurboQuant: 27x compression at 99.8% recall@10 (identical reranking). Larger-than-RAM sharded/memmap search, tqp CLI with distribution-free rank certificates, asymmetric K/V, CUDA+Triton kernels, auto-config. 919 tests. MIT.

Python

Differential KV compression: delta-encode adjacent token KV before PolarQuant 4 months ago

enhancement help wanted

ahb-sjsu/turboquant-pro

22

Embedding & KV-cache compression for LLMs and vector databases. PCA-Matryoshka + TurboQuant: 27x compression at 99.8% recall@10 (identical reranking). Larger-than-RAM sharded/memmap search, tqp CLI with distribution-free rank certificates, asymmetric K/V, CUDA+Triton kernels, auto-config. 919 tests. MIT.

Python

Add L2 (Euclidean) distance function and <-> operator 4 months ago

enhancement good first issue

ahb-sjsu/turboquant-pro

22

Embedding & KV-cache compression for LLMs and vector databases. PCA-Matryoshka + TurboQuant: 27x compression at 99.8% recall@10 (identical reranking). Larger-than-RAM sharded/memmap search, tqp CLI with distribution-free rank certificates, asymmetric K/V, CUDA+Triton kernels, auto-config. 919 tests. MIT.

Python

Open Issues Need Help