←all projects
Vesta
privateSystemsZero-dependency C++ CPU inference engine for ternary neural networks ({-1, 0, +1} weights). Bit-packs 64 weights into dual 64-bit bit-planes and executes branchless SIMD kernels hitting up to 30.9 Gop/s on Apple Silicon without GPUs, PyTorch, or BLAS.
Highlights
- Custom VESTABT1 binary format packs 64 weights into dual 64-bit masks (+1/-1), computed via branchless SIMD matvec kernels with zero CPU mispredictions in the hot loop
- Hand-vectorized ARM NEON, AVX2, AVX-512, and scalar backends delivering 10.96 Gop/s single-core and 30.9 Gop/s multi-core on Apple M4, all checksum-verified for bit-exact parity
- Full autoregressive transformer decode with RMSNorm, zero-transcendental RoPE lookups, a preallocated reusable KV cache, and causal multi-head attention
- Self-contained end-to-end toolchain featuring a standard-library Python trainer and error-compensated quantizer that trains and exports directly to .bits models
Stack
- ARM NEON
- AVX-512
- AVX2
- Bit Manipulation
- C++
- Quantization
- SIMD
Languages
- Python 53%
- C++ 41%
- Shell 5%
- Makefile 1%
Activity
60 commits · updated last month
Components
ML Systems & High-Performance Computing
Zero-dependency C++ CPU inference engine for ternary neural networks ({-1, 0, +1} weights). Bit-packs 64 weights into dual 64-bit bit-planes and executes branchless SIMD kernels hitting up to 30.9 Gop/s on Apple Silicon without GPUs, PyTorch, or BLAS.
- C++
- ARM NEON
- AVX2
- AVX-512
- SIMD
- Quantization
- Bit Manipulation