Crate rlx_voxtral_tts 

Voxtral-4B-TTS on RLX — Ministral LM + acoustic flow matching + codec decode.

Native Rust port of vLLM-Omni VoxtralTTSAudioGeneration (no Python at inference).

Re-exports

pub use backbone::CompiledMinistralLm;

pub use backbone::MinistralLm;

pub use backbone::NativeTtsEngine;

pub use bench::VoxtralTtsBenchReport;

pub use codec::CodecDecoder;

pub use config::HF_MODEL_ID;

pub use config::VoxtralTtsConfig;

pub use generation::GenerationConfig;

pub use load::VoxtralTtsWeightStore;

pub use lora::load_lora_bank;

pub use options::VoxtralTtsOptions;

pub use options::VoxtralTtsRunnerBuilder;

pub use prompt_tokens::load_prompt_tokens;

pub use runner::VoxtralTtsRunner;

pub use runner::parse_codes_file;

pub use runner::write_wav_mono;

pub use tokens::PRESET_VOICES;

pub use voice::VoiceEmbedding;

pub use voice_clone::VoiceCloneSupport;

pub use voice_clone::clone_from_reference_audio;

pub use voice_clone::encode_reference_wav;

pub use voice_clone::encode_reference_wav_to_file;

pub use voice_clone::max_reference_seconds;

pub use voice_clone::voice_clone_support;

Modules

acoustic

Flow-matching acoustic transformer (vLLM-Omni FlowMatchingAudioTransformer).

acoustic_compiled

Compiled acoustic velocity stack (3-token FM sequence, bidirectional attention).

acoustic_engine

Acoustic head backend — eager CPU reference or RLX-compiled stack.

acoustic_flow

Compiled acoustic velocity stack (3-token bidirectional transformer, no attention RoPE).

backbone

bench

Stage timing for native TTS (LM prefill / decode, acoustic, codec).

cli

codec

config

Voxtral-4B-TTS config (params.json / HF layout).

decode_shard_layer

Decode layer for wgpu LM shards — global checkpoint keys, local past_k_* inputs.

generation

Generation options (parity with vLLM-Omni sampling).

lm_flow

Compiled Ministral graphs (inputs_embeds prefill/decode, no LM head).

load

Mmap-backed weight access for consolidated.safetensors.

lora

LoRA adapters on Ministral attention + FFN projections (inference merge + eager apply).

math

Small ndarray helpers for eager CPU inference.

options

Runner options — device and eager fallbacks.

prompt_tokens

Load prompt token ids exported by the Docker tools image.

rng

Reproducible Gaussian noise for flow-matching (seeded PCG).

runner

End-to-end TTS runner — native Rust only.

speech_tokenizer

Native Tekken speech prompt tokenization (replaces Docker mistral_common).

tokens

Audio + text special tokens (vLLM-Omni AudioSpecialTokens).

voice

Preset voice embeddings (voice_embedding/*.pt or converted .f32).

voice_clone

Reference-audio voice cloning via trained codec encoder weights.

voice_pt

Convert HuggingFace voice_embedding/*.pt (bf16 zip) to native .f32.

weights

Map Voxtral-4B-TTS checkpoint keys → Llama builder keys.