Frosty40/YuE2-Turbo-Radio

YuE2 Turbo Radio

Measured real-time music generation with YuE2 3B, now including Apollo's RTX 4080 CUDA port. This repository now contains the actual Q8_0 model, F16 VAE, tokenizer and config files, plus the optimized source/runtime release and listening gallery. It is not a new fine-tune. It uses m-a-p/YuE2-3B via audio-cpp/Yue2-3B-GGUF and audio.cpp. The exact benchmarked model files are hosted here, unchanged from the attributed upstream GGUF release.

Download the actual model

The model is hosted in this repository—not just linked elsewhere. The model bundle is 4.53 GB total (decimal), including the exact files used in the published benchmarks:

File	Download	Size
YuE2 3B Q8_0	Actual GGUF weights	4.264 GB
YuE2 F16 VAE	Actual VAE weights	265.2 MB
Tokenizer + configs	Browse the complete bundle	2.6 MB

pip install -U huggingface_hub
# Model + VAE + sidecars only:
hf download Frosty40/YuE2-Turbo-Radio --local-dir yue2-turbo-radio \
  --include 'models/Yue2-3B-GGUF/*'
# Or get the full model/runtime/source package:
hf download Frosty40/YuE2-Turbo-Radio --local-dir yue2-turbo-radio
cd yue2-turbo-radio

Model provenance and SHA256s · CC BY-NC 4.0 license · Build/run instructions. Weights are the unchanged audio-cpp conversion of m-a-p/YuE2-3B—not an original fine-tune. The measured speedups require the included optimized runtime; copying the weights alone does not accelerate stock software.

Listen: Knight Shift Radio

Tune in to Knight Shift Radio →

Original YuE2 music, with warm Kasey song introductions and separate Jessica tech/AI information segments. The show's direction is developed arrangements, thoughtful genre transitions and positive lyrics about kindness, hard work and shared progress. Speech uses ElevenLabs separately from YuE2 music. The public site is linked here; stream availability and the live rundown can change. No claim is made that the samples below are currently scheduled or on air.

Latest measured results — Apollo RTX 4080, 2026-11-02

RTX 4080 16 GB; CUDA 13.0, SM 8.9; Q8_0 YuE2-3B + F16 VAE. GPU clocks and power limits unchanged. One warmup discarded, then three separate measured cold CLI processes per full-song configuration. Speed is audio seconds / elapsed seconds. Session and startup-inclusive numbers are deliberately separate.

Workload	NAR steps	Audio	Median session speed	Median including CLI startup
city_lights, cache-off baseline	8	69.00 s	3.173×	3.030×
city_lights, accepted full-song runtime	8	69.00 s	3.168×	3.019×
Same accepted runtime	32	69.00 s	2.754×	2.648×
Four-minute song	8	257.08 s	3.322×	3.269×

In practical terms, the 4:17 song takes about 79 seconds including CLI startup. At 32 steps, city_lights takes about 26 seconds including startup for 69 seconds of audio. The cache-on/off difference is within noise; no CUDA prefix-cache speedup is claimed. Long-song NAR takes 10.09–10.19 s versus about 3.8–4.0 s on the short song: no superlinear penalty was observed at the measured window size. This is not a universal scaling law.

• Short-song runs · Long-song runs
• Acceptance summary · Full results and rejected experiments
• Build/reconstruction instructions · Source/runtime manifests

Streaming: first audio in 8.56–9.43 seconds, 3/3 passed

The adapter-safe asynchronous NAR preload runtime produces first audio including cold CLI startup at 8.555, 9.165 and 9.432 seconds. Each run naturally completes 256.839 s of stereo PCM in 21 contiguous chunks, with zero late chunks and zero gaps. Its PCM is byte-identical to the original streaming predecessor for this exact request. This does not claim bit parity between CUDA and SYCL, model-quality parity, or tested LoRA output parity (adapter sources fall back to the ordinary loader).

The original, pinned-upload, retained-mapping and earlier-score variants that missed the latency gate remain in the receipts. They are not promoted or silently hidden. All three accepted runs · Accepted streaming identity.

Thirty-minute playlist: passed

1,800 seconds of real-paced PCM consumption, 86,400,000 stereo frames, zero queue failures, zero fallback items and zero inserted-silence frames; expected/actual PCM SHA256 match. This is an isolated producer/consumer soak, not a network-stream or physical-audio-device test, and does not establish listener delivery or perfect musical transitions. Playlist receipt.

New listening gallery — 16 original 32-step songs

51 minutes 35 seconds, generated serially on Apollo: four house, four trap, four wub/bass and four industrial-ambient tracks. Written lyrics emphasize kindness, patient work and shared progress. The arrangements request evolving sections, instrumental interplay and a natural ending. Completion/non-silence checks passed; musical coherence, genre fit and vocal intelligibility still require human listening. These are samples, not a controlled 16-song speed benchmark. MP3s are unnormalized; there is no limiter hiding PCM rail counts.

Complete lyrics and prompts · Per-song duration, render wall time, hashes and signal diagnostics.

#	Genre	Song	Duration	Render wall	Realtime
1	house	Open Door Frequency	4:01	85.9 s	2.81×
2	house	Brick by Brick	2:06	46.2 s	2.73×
3	house	Common Ground	3:20	73.4 s	2.72×
4	house	Carry the Dawn	4:32	98.0 s	2.78×
5	trap	Hands to the Work	2:06	46.5 s	2.72×
6	trap	Small Wins Skyline	2:01	46.2 s	2.62×
7	trap	Make Room	2:43	55.2 s	2.96×
8	trap	Rise Together	2:05	43.0 s	2.91×
9	wub	Kindness Resonance	3:14	70.7 s	2.74×
10	wub	Soft Power	3:55	77.4 s	3.03×
11	wub	Lift the Signal	3:18	69.7 s	2.84×
12	wub	Low Tide Lanterns	3:46	83.1 s	2.72×
13	industrial-ambient	Shift Change Stars	3:34	74.6 s	2.86×
14	industrial-ambient	Warmth in the Machine	4:16	87.8 s	2.92×
15	industrial-ambient	Rust Gardens	3:40	73.3 s	3.00×
16	industrial-ambient	The Lights We Leave	2:58	57.0 s	3.12×

Benchmark listening samples

• city_lights — 8 steps
• city_lights — same runtime, 32 steps
• Four-minute song — 8 steps
• Source-WAV hashes and selected-run metadata

B70 results and research — history retained, latest recheck disclosed

The previously published November 1 B70 measurements remain available: 3.091× session realtime on city_lights at 8 steps, 2.414× at 32 steps, 2.141× on Knight Won't Stop at 32 steps, first streaming audio 7.58–7.73 s, and a passing 30-minute isolated queue. Previous card and linked receipts.

The November 2 canonical B70 recheck failed the 3.0× gate, median 2.818× (2.818, 2.141, 2.939×). All three songs naturally completed and exceeded 55 seconds. This does not erase the earlier pass and is not an Apollo result. Latest B70 recheck receipt. No new B70 GPU execution was performed for this publication.

Recent research/audits cover native sampler behavior, seed contract, generation defaults, Q8/F16 cache and PCM rail accounting, VAE tiling, decoder precision and official-decoder comparisons. These are scoped experimental or diagnostic findings, not all production improvements or end-to-end speedups. Research index · Evidence reconciliation. The historical October recipe, charts and earlier songs remain in this repository; read the prior card for that material.

What changed on Apollo

The CUDA port carries over growing AR buffers, cached NAR prefixes, bounded NAR windows (1024 frames / 128 overlap), and deferred packed-weight uploads from the B70 work. CUDA Flash Attention and CUDA graphs replace Intel-specific kernels. An independent, opt-in streaming pipeline overlaps safe NAR weight loading with earlier work and preserves the tested predecessor PCM. The measured full-song runtime stays frozen; streaming has its own binary and source identity. The older engine ignores yue2.packed_projections and yue2.kv_cache_alignment request keys: they are retained where present in raw historical receipts, but receive no performance credit.

Reproduce / download

Start with Apollo build and run instructions. The pinned source base, patches, source archive SHA256, accepted executable SHA256s, model hashes and protocol receipts are included. Linux x86-64 CUDA executables are provided as experimental, hardware/environment-specific artifacts, not a portable installer. Rebuilding is supported by the recorded source recipe, but byte-identical rebuilds are not claimed. Download the model and runtime together from this repo; the model remains CC BY-NC 4.0. Nothing in this release changes your radio queue.

Limitations, attribution and licenses

These are local measured results, not official benchmark/leaderboard scores or a world-record claim. No official-model waveform/quality parity is established. Cross-hardware sampled trajectories can differ. Signal checks are not listening approval: some takes have PCM rail samples and naturally quiet tails. The queue test does not certify internet delivery, actual loudspeaker output, host cadence or genre transitions.

• YuE2 3B / VAE: Multimodal Art Projection (m-a-p), CC BY-NC 4.0; GGUF conversion by audio-cpp.
• audio.cpp: ShugoAI LLC, Apache-2.0. Its bundled third-party components retain their licenses; the source archive includes those notices. SYCL EXPM1/concat work derives from ggml/llama.cpp (MIT).
• Project engine patches and scripts: Apache-2.0; model card, charts and generated music: CC BY-NC 4.0.
• The previous card preserves the detailed contributor/tool credits and historical release narrative. NOTICE records the existing licensing credits. Not affiliated with or endorsed by m-a-p, ShugoAI, NVIDIA, Intel, or the ggml authors.