kaio 0.5.0

Rust-native GPU kernel authoring framework. Write GPU compute kernels in Rust, automatically lower to PTX. Cross-platform (Windows + Linux), type-safe, no CUDA C++ required.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41

//! # KAIO
//!
//! **Rust-native GPU kernel authoring framework.**
//!
//! KAIO (καίω — to kindle, to ignite) lets developers write GPU compute
//! kernels in Rust and lower them to PTX for execution on NVIDIA GPUs.
//! A Rust alternative to OpenAI's Triton, targeting Windows and Linux
//! from day one, with automatic PTX generation and Rust's type-safety
//! guarantees.
//!
//! ## Crates
//!
//! - [`kaio_core`] — PTX IR types, instruction emitters, PtxWriter
//! - [`kaio_runtime`] — CUDA device management, buffers, PTX loading, kernel launch
//! - `kaio_macros` — `#[gpu_kernel]` proc macro (re-exported here)
//! - `kaio_ops` — pre-built GPU operations (scalar + tensor-core matmul
//!   in f16/bf16, quantized INT8/INT4 matmul, fused QKV projections,
//!   attention and FlashAttention with backward, auto-tuners).
//!   Separate crate, not re-exported from `kaio`. Add with `cargo add kaio-ops`.
//!
//! ## Status
//!
//! **Phase 9 complete (v0.5.0).** Tensor-core matmul family (f16 +
//! bf16, sync + `cp.async`), quantized INT8/INT4 ops, FlashAttention
//! forward + backward, candle bridge (`kaio-candle`), `ldmatrix` IR
//! primitive. See the repository README for the full feature table
//! and roadmap.

#![warn(missing_docs)]

pub mod gpu_builtins;
pub mod prelude;

/// PTX code generation — IR types, instruction emitters, and PTX text emission.
pub use kaio_core as core;

/// CUDA runtime — device management, buffers, PTX loading, kernel launch.
pub use kaio_runtime as runtime;

/// Re-export the `#[gpu_kernel]` attribute macro.
pub use kaio_macros::gpu_kernel;