Crate rlx_nemotron 

NVIDIA Nemotron 3 Nano runner.

Nemotron ships as several GGUF arch tags:

• nemotron — text-only, Llama-shaped attention stack; runs via the rlx_llama32::Llama32Runner delegate below.
• nemotron_h / nemotron_h_moe — hybrid Mamba2 + attention; the NemotronHybridRunner in runner.rs drives it via per-layer Mamba2StepStage interleaved with stateless attention blocks.

The Omni 30B variant (vision + audio) lives in rlx-nemotron-omni and is wired independently.

Re-exports

pub use config::NemotronHybridConfig;

pub use config::NemotronLayerKind;

pub use flow::mamba2_decode_layer_plugin_with_sink;

pub use flow::stateless_attention_layer_plugin;

pub use runner::NemotronHybridRunner;

pub use runner::NemotronHybridRunnerBuilder;

Modules

config

Nemotron-H hybrid Mamba+attention configuration.

flow

Nemotron-H per-layer decode-step blocks.

runner

Nemotron-H hybrid runner — single-step decode with Mamba2 state.

Structs

Llama32Runner

Llama32RunnerBuilder

NemotronRunner

NemotronRunnerBuilder

Constants

FAMILY

PLAN_MILESTONE

Functions

cli_run

Type Aliases

Llama32ConfigSource

Where to load the Llama 3.2 config from.