Crate eegdino_rs 

eegdino-rs

Rust inference crate for the EEG-DINO foundation model, built on RLX.

EEG-DINO learns robust EEG representations via hierarchical self-distillation on 9 000+ hours of EEG data. This crate provides a faithful port of the encoder architecture with verified numerical parity (NRMSE < 1e-6) against the original PyTorch implementation on CPU, Metal, and MLX backends.

Model sizes

Variant	Params	d_model	Heads	Layers	FFN dim
Small	4.6 M	200	8	12	512
Medium	33 M	512	16	16	1 024
Large	201 M	1 024	16	24	2 048

Quick start

use eegdino_rs::prelude::*;

let device = parse_device("metal")?; // cpu | metal | mlx | gpu
let (mut encoder, load_ms) = EegDinoEncoder::load(
    "weights/eeg_dino_small.safetensors".as_ref(),
    None,
    device,
)?;

let signal = vec![0.0f32; 19 * 2000];
let result = encoder.encode_raw(&signal, 1, 19, 2000)?;
// result.shape == [1, 191, 200]

Backends

Feature / device	Backend	Notes
cpu, rlx-cpu	RLX CPU	Rayon + SIMD; default
metal	Apple Metal / MPS	macOS
mlx	Apple MLX	macOS
gpu, wgpu	RLX wgpu	Metal/Vulkan/DX12 (parity vs CPU in progress)

Enable all with --features all-backends.

Re-exports

pub use config::ModelConfig;

pub use config::ModelSize;

pub use error::EegDinoError;

pub use error::Result;

pub use rlx::detect_model_size;

pub use rlx::device_label;

pub use rlx::feature_for;

pub use rlx::is_device_available;

pub use rlx::parse_device;

pub use rlx::EegDinoEncoder;

pub use rlx::EegDinoEncoderBuilder;

pub use rlx::EncodingResult;

Modules

config

error

Typed error type for the eegdino-rs public API.

prelude

Convenience re-exports for common usage patterns.

rlx

RLX-backed EEG-DINO inference (rlx::Graph + rlx::Session).

Functions

init_threads

Configure the Rayon thread pool. Call once before model use.