ramvamp-core 0.1.1

Runtime for streaming fine-grained MoE experts from NVMe: packed format, io_uring reads, LFU expert cache, CPU kernels, KV cache, generation loop
Documentation
//! Compute kernels behind a backend trait.
//!
//! The first backend is CPU: runtime-detected AVX2 paths for quantized GEMV,
//! grouped MoE FFN (fused dequant + activation + weighted reduction), RMSNorm,
//! RoPE, attention, and sampling. A Vulkan backend is planned behind the same
//! trait; keeping the boundary here from day one is deliberate.
//!
//! Layout rule learned the hard way upstream: packed sub-tensor offsets may
//! guarantee less alignment than a wide load assumes. Every vectorized path
//! must state, and test, the alignment it relies on.

pub mod attention;
pub mod backend;
mod error;
pub mod gemv;

pub mod quants;

pub mod primitives;

pub use backend::{Backend, CpuBackend};
pub use error::KernelError;
pub use gemv::{
    gemv_q8_0, gemv_q8_0_batched, gemv_q8_0_rows, gemv_q8_k, gemv_q8_k_batched, gemv_q8_k_rows,
};