Aurora-8B-Ultra-RL: 100% Pure-Rust Burn + CubeCL p/q Quantal Full-Duplex Audio-LM (LFM2.5-8B-A1B + Kokoro-82M + ESP32-S3 + Discord)
  • Rust 97.5%
  • Shell 1.7%
  • Just 0.8%
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
2026-11-02 04:15:10 +09:00
.cargo feat(aurora-8b): complete pure-Rust Burn + CubeCL p/q quantal audio-LM monorepo 2026-11-01 20:30:17 +09:00
.github/workflows feat(e2e): add zero-Python Safetensors parser, MEMS/RIR acoustic augmenter, WAV exporter & GitHub CI 2026-11-02 04:07:04 +09:00
crates feat(gateway,weights): add OpenAI /v1/audio/speech WAV/PCM endpoint, /v1/realtime WS & Safetensors CLI 2026-11-02 04:15:10 +09:00
.gitignore feat(aurora-8b): complete pure-Rust Burn + CubeCL p/q quantal audio-LM monorepo 2026-11-01 20:30:17 +09:00
bootstrap_aurora.sh feat(aurora-8b): complete pure-Rust Burn + CubeCL p/q quantal audio-LM monorepo 2026-11-01 20:30:17 +09:00
Cargo.lock feat(train-rl): add Omni one-shot E2E training prompt & 5-stage AuroraE2eTrainer pipeline 2026-11-01 21:25:33 +09:00
Cargo.toml feat(aurora-8b): complete pure-Rust Burn + CubeCL p/q quantal audio-LM monorepo 2026-11-01 20:30:17 +09:00
Justfile feat(aurora-8b): complete pure-Rust Burn + CubeCL p/q quantal audio-LM monorepo 2026-11-01 20:30:17 +09:00
opencode.json feat(aurora-8b): complete pure-Rust Burn + CubeCL p/q quantal audio-LM monorepo 2026-11-01 20:30:17 +09:00
PROMPT_TRAIN_E2E_ULTRAOMNI.md feat(train-rl): add Omni one-shot E2E training prompt & 5-stage AuroraE2eTrainer pipeline 2026-11-01 21:25:33 +09:00
PROMPT_ULTRAOMNI.md feat(aurora-8b): complete pure-Rust Burn + CubeCL p/q quantal audio-LM monorepo 2026-11-01 20:30:17 +09:00
README.md feat(aurora-8b): complete pure-Rust Burn + CubeCL p/q quantal audio-LM monorepo 2026-11-01 20:30:17 +09:00
rust-toolchain.toml feat(aurora-8b): complete pure-Rust Burn + CubeCL p/q quantal audio-LM monorepo 2026-11-01 20:30:17 +09:00

Aurora-8B-Ultra-RL (aurora-8b-ultra-rl)

100% Pure-Rust End-to-End p/q Quantal Full-Duplex Audio-Language Model built natively on Burn (0.22.1) and CubeCL (0.5+) for ESP32-S3 / ESP32-P4 bare-metal edge nodes and Discord Full-Duplex Voice Bots.


1. Unified Model Synergy (Aurora-8B-Liquid-Kokoro)

Aurora-8B-Ultra-RL fuses four state-of-the-art open-weights architectures into a single pure-Rust Burn computation graph:

Subsystem Source Checkpoints Architectural Role in Aurora-8B-Ultra-RL
8B Hybrid MoE Backbone LiquidAI/LFM2.5-8B-A1B 8.3B total / 1.5B active parameters across 24 layers: 18 Double-Gated ShortConv layers (conv_L_cache = 3) + 6 GQA layers (32 Q / 8 KV heads, QK-RMSNorm, \theta = 10^6) at [2, 6, 10, 14, 18, 21], with 2 dense SwiGLU layers (0..2) and 22 Sparse 32-Expert MoE layers (top-4, normalized sigmoid affinity routing).
RL Teacher & Speculative MTP XiaomiMiMo/MiMo-V2.6-Pro-RL Provides the 5-layer Multi-Token Prediction (MTP) head drafting 5 future tokens per step, plus MOPD2 (Multi-Teacher On-Policy Distillation) and GRPO (Group-Relative Policy Optimization).
Streaming ASR, 80ms Clock & 8-Class VAD Edge0/Audio8-ASR-Infinite + LiquidAI/LFM2.5-Audio-1.5B 8-Codebook RVQ Depthformer (q=8, dim=256, layers=6, codes 0..2047, stop 2048) and 8-Class Multi-Horizon Semantic VAD (Silence, SpeechActive, ThinkingPause, EndOfTurn_0_25s, EndOfTurn_0_5s, EndOfTurn_1_0s, Backchannel, BargeInInterrupt).
Expressive TTS Synthesis hexgrad/Kokoro-82M + Edge0/Audio8-TTS-Realistic-0.6b Native Burn Kokoro-82M (StyleTTS2 + ISTFTNet 24 kHz generator with AdaIN1d, Snake1d, 9-harmonic SourceModuleHnNSF with Hillis-Steele O(\log T) parallel scan, and exact COLA conv_transpose1d synthesis).
Bare-Metal ESP32-S3 Edge Node Edge0/Audio8-ASR-0.1B (Distilled) #![no_std] burn-flex INT8 depthwise-separable micro-encoder + 2-codebook RVQ quantizer + NLMS Acoustic Echo Cancellation (AEC) emitting 42-byte Quantal80msPacket frames (4.2 kbps).

2. The p/q Quantal Objective

Continuous speech is discretized at an 80 ms (12.5 Hz) frame clock into p=1 semantic audio token and q=8 acoustic RVQ codebooks alongside an 8-class Semantic VAD distribution:

\mathcal{L}{\text{Aurora}} = \sum{t=1}^{T} \left[ \alpha \log P_{\theta}(s_t \mid \mathbf{h}t) + \beta \sum{j=1}^{8} \log P_{\theta}(a_{t,j} \mid \mathbf{h}t, s_t, a{t,<j}) + \gamma \mathcal{L}_{\text{VAD}}(\mathbf{v}t \mid \mathbf{h}t) + \lambda{\text{STFT}} \mathcal{L}{\text{MR-STFT}}(\hat{y}, y) \right]


3. Repository Structure

aurora-8b-ultra-rl/
├── Cargo.toml                      # Workspace root (9 crates)
├── Justfile                        # Build, test, weight export, RL train, ESP32 flash & multi-remote push
├── bootstrap_aurora.sh             # Multi-remote GitHub / Codeberg / g8t & OpenCode DeepSeek-V4 bootstrap
├── opencode.json                   # OpenCode configuration for deepseek/deepseek-v4
├── PROMPT_ULTRAOMNI.md             # Master One-Shot UltraOmni architectural specification
└── crates/
    ├── aurora-core/                # #[no_std] Quantal80msPacket (42B), NLMS AEC, polyphase FIR resamplers
    ├── aurora-weights/             # Safetensors -> Burnpack (.bpk) converter, LFM2.5-8B-A1B & Kokoro-82M modules
    ├── aurora-kernels/             # Fused LFM2.5 ShortConv, 8-codebook Depthformer & Autodiff backward bridge
    ├── aurora-kokoro-tts/          # Pure-Rust Burn Kokoro-82M ISTFTNet (N_FFT=20, hop=5, win=20) + MR-STFT loss
    ├── aurora-train-rl/            # 5-stage RL pipeline (Projector -> SFT -> MOPD2 -> GRPO -> Quant) + 5-layer MTP
    ├── aurora-discord/             # Serenity + Songbird full-duplex RTP voice engine + <1ms barge-in flush
    ├── aurora-esp32-gateway/       # Axum /v1/aurora-esp32 WebSocket server + 960B (30ms) RFC-6455 chunk pacer
    ├── aurora-unified-daemon/      # Single-process GPU daemon sharing 5.2GB VRAM across Discord & ESP32 streams
    └── aurora-esp32-node/          # Bare-metal ESP32-S3 (#![no_std] esp-hal + burn-flex) firmware & WS parser

4. Quick Start

# 1. Type-check and run all unit & architectural invariant tests across the workspace
cargo test --workspace

# 2. Run the 5-stage RL training pipeline (Projector -> SFT -> MOPD2 -> GRPO -> Quant)
cargo run --release -p aurora-train-rl -- --stage all

# 3. Launch the unified full-duplex GPU daemon (Discord Voice + Axum ESP32 WebSocket Gateway)
DISCORD_TOKEN="your_bot_token" \
ESP32_WS_BIND="0.0.0.0:8080" \
cargo run --release -p aurora-unified-daemon