Yue2 support, waiting for pull request

[yurii-sio2]

I'm super waiting for Yue2 to be merged into main branch! 🙏🏽
Thanks for your hard work! 🙌

If you can build and make binary for ubuntu-x64-vulkan, it whould be great, I'd test it on my setup (AMD Strix Halo).

[0xShug0]

@yurii-sio2 Building now! Once finished, you can find them in the “Artifacts” section on this page: https://github.com/0xShug0/audio.cpp/actions/runs/34642627647

[yurii-sio2]

Thank you. Tested on my hardware and it works perfectly! I have Ubuntu 26, AMD Ryzen AI Max 395+, 96 GB VRAM.
Generation of 3 minutes song took 5:14 minutes running Vulkan version.
If needed any other stats, ask me.

To make it work I had to download all files from 'sidecars' folder and place the files at the 'sidecars' folder near the 'audiocpp_cli' file
https://huggingface.co/audio-cpp/Yue2-3B-GGUF/tree/main/sidecars

- audio.cpp
---- audiocpp_cli
---- sidecars
-------- yue2-vae-config.json
-------- ... other json and tiktoken files

My run command with Vulkan:

./audiocpp_cli \
  --task gen \
  --family yue2 \
  --model yue2-3b-bf16.gguf \
  --backend vulkan \
  --threads 8 \
  --text $'[Verse]\nText here\n[Chorus]\nText here' \
  --request-option 'style=English, piano pop...' \
  --request-option cot=full \
  --request-option seed=20260912 \
  --request-option num_inference_steps=10\
  --session-option yue2.model_gguf=yue2-3b-bf16.gguf \
  --session-option yue2.vae_gguf=yue2-vae-f32.gguf \
  --out demo2.wav \
  --log

[0xShug0]

Thanks for testing!

[IIIIIllllIIIIIlllll]

released today

[yurii-sio2]

For those, who interested in more finetuned params for ABC and others, here is the way to run it.

./audiocpp_cli \
  --task gen --family yue2 --model yue2-3b-bf16.gguf --backend vulkan --threads 8 \
  --text $'[intro 0:00-0:30]\n[verse 0:30-1:20]\n[chorus 1:20-2:10]\n[verse 2:10-3:00]\n[chorus 3:00-3:50]\n[bridge 3:50-4:30]\n[outro 4:30-5:00]\n' \
  --request-option 'style=instrumental, dark phonk, heavy 808 bass...' \
  --request-option guidance_scale=1.1 \
  --request-option cot=full --request-option num_inference_steps=10 \
  --request-option seed="$(shuf -i 1-2147483647 -n 1)" \
  --request-option nar_noise=generated \
  --request-option abc_temperature=0.9 \
  --request-option abc_top_p=0.9 \
  --request-option abc_top_k=30 \
  --request-option abc_repetition_penalty=0.956 \
  --request-option abc_penalty_window=200 \
  --request-option abc_min_tokens=1000 \
  --request-option abc_max_tokens=6000 \
  --request-option semantic_temperature=0.9 \
  --request-option semantic_top_p=0.95 \
  --request-option semantic_top_k=100 \
  --request-option semantic_repetition_penalty=1.11 \
  --request-option semantic_penalty_window=100 \
  --request-option semantic_min_tokens=2000 \
  --request-option semantic_max_tokens=8000 \
  --session-option yue2.model_gguf=yue2-3b-bf16.gguf --session-option yue2.vae_gguf=yue2-vae-f32.gguf \
  --out "$(date +%Y-%m-%d_%H-%M-%S).wav" \
  --metrics --log

If you want to provide your own ABC, then add parameter:
--request-option abc=<your ABC here>

[0xShug0]

@yurii-sio2 Thanks!

@mlsterpr0 maybe useful for you.

[0xShug0]

BTW @yurii-sio2 @IIIIIllllIIIIIlllll I talked with an AI Applications Engineer at AMD about the performance issue on the Ryzen AI Max+ 395, which looks pretty slow to me for such a high-end GPU. He suggested Hyperloom for AMD optimizations. You may want to take a look at it.