mirror of
https://github.com/hanzoai/papers.git
synced 2026-10-22 17:20:17 +00:00
18 lines
1.1 KiB
TeX
18 lines
1.1 KiB
TeX
\section{Background}
|
|
|
|
|
|
\subsection{Rust for ML}
|
|
|
|
Rust provides memory safety guarantees without garbage collection, making it ideal for performance-critical ML code. The ownership model prevents data races in parallel computation, while zero-cost abstractions enable high-level APIs without runtime overhead.
|
|
|
|
Prior work on Rust ML frameworks includes tch-rs (PyTorch bindings)~\cite{tchrs2020} and burn~\cite{burn2023}. Our framework differs by prioritizing deployment efficiency and FHE compatibility over training performance.
|
|
|
|
\subsection{Fully Homomorphic Encryption}
|
|
|
|
FHE schemes like TFHE~\cite{tfhe2016} and CKKS~\cite{ckks2017} enable computation on encrypted data. However, FHE supports only polynomial operations efficiently; non-polynomial functions (ReLU, sigmoid, softmax) require approximation or bootstrapping.
|
|
|
|
\subsection{Quantization}
|
|
|
|
Model quantization reduces precision from FP32 to INT8/INT4, decreasing memory bandwidth and enabling vectorized computation. The GGML format~\cite{ggml2023} introduces k-quants (Q4\_K, Q5\_K, Q6\_K) that combine multiple quantization scales per block for improved accuracy.
|
|
|