Crate lumamba 

lumamba-rs — LuMamba EEG foundation-model inference in Rust

Pure-Rust inference for LuMamba (PulpBio/LuMamba, from BioFoundation), built on the RLX compiler/runtime.

LuMamba reuses LUNA’s topology-invariant front-end — variable-channel EEG is compressed into a fixed set of learned queries by cross-attention — and replaces LUNA’s rotary Transformer encoder with a stack of bidirectional Mamba (FEMBA) blocks that run in linear time over the patch sequence.

EEG (B, C, T)
  ├─ PatchEmbed (3× Conv2d + GroupNorm + GELU)         ┐  CPU prepare
  └─ FreqEmbed  (FFT magnitude/phase → MLP)            ├─ (host)
             → + NeRF channel-location embedding       ┘
  → CrossAttention channel unification (C → Q queries)  ┐
  → reshape [B, S, Q·E]                                 │  RLX graph
  → N × { LayerNorm → BiMamba(fwd + flip∘rev) → +res }  │
  → reconstruction head  →  (B, C, T)                   ┘

The temporal recurrence runs on RLX’s first-class selective_scan op (Mamba-1 SSM), the depthwise causal conv on grouped conv2d, and everything else on the standard graph ops — so the whole model compiles to a single RLX CompiledGraph per input shape and runs on any RLX backend (cpu, metal, mlx, cuda, …).

Re-exports

pub use config::ModelConfig;

pub use rlx::ClassifierKind;

pub use rlx::EpochEmbedding;

pub use rlx::LuMambaEncoder;

pub use rlx::RlxEpoch;

pub use rlx::RunEpochOpts;

pub use channel_positions::bipolar_channel_xyz;

pub use channel_positions::channel_xyz;

pub use channel_positions::montage_channels;

pub use channel_positions::nearest_channel;

pub use channel_positions::normalise;

pub use channel_positions::MontageLayout;

pub use channel_vocab::channel_index;

pub use channel_vocab::channel_indices;

pub use channel_vocab::channel_indices_unwrap;

pub use channel_vocab::CHANNEL_VOCAB;

pub use channel_vocab::SEED_CHANNELS;

pub use channel_vocab::SIENA_CHANNELS;

pub use channel_vocab::TUEG_CHANNELS;

pub use channel_vocab::VOCAB_SIZE;

Modules

channel_positions

EEG channel position lookup from embedded standard montage files.

channel_vocab

Global channel name vocabulary for LUNA.

config

Model and runtime configuration for LuMamba inference.

eval

Downstream evaluation harness: load a labeled EEG eval set, run the fine-tuned LuMamba classifier per epoch, optionally aggregate epoch predictions to the recording level, and score with the paper’s metrics.

hf

HuggingFace Hub weight resolution.

metrics

Classification metrics for downstream evaluation.

rlx

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

Structs

LuMambaBackends

BackendSupport for the LuMamba model family — the RLX 0.2.10 idiom for declaring which devices a model can execute on, so device validation yields a uniform error instead of a hand-rolled match ladder.

Constants

SUPPORTED_DEVICES

The RLX backends lumamba-rs forwards Cargo features for. A --device token outside this set is rejected at parse time by parse_device.

Functions

init_threads

Configure the global Rayon thread pool (used by the CPU prepare path and the RLX CPU backend).

parse_device

Parse a --device CLI token into an rlx::Device.