max / everycycle

489 B · 17 lines History Blame Raw

1	#![allow(clippy::doc_markdown)]
2
3	//! EveryCycle inference runtime.
4	//!
5	//! Loads quantized models, places layers across heterogeneous compute (GPU,
6	//! CPU, NVMe overflow), drives the forward pass, and streams tokens to the
7	//! scheduler.
8	//!
9	//! See `docs/architecture.md` in the workspace root for the broader sketch.
10
11	pub mod executor;
12
13	pub use executor::{
14	ActivationHandle, BoundDevice, Dtype, Executor, ExecutorError, PortableActivation,
15	SubGraphPlan, TensorShape, Vendor,
16	};
17