connect core.py with zluda on applio

[jaunty flame]

so, the problem here is that my gpu is amd, well, igpu to be corrected: 780m. I install applio in my windows laptop and setup zluda succesfully. i tried to run with core.py cli so that i can automate some tasks, but it run on cpu instead. Now i tried to find a way to run zluda inside cli core.py, or something similar as long as i can use my amd gpu while connect to applio through api. As for via gradio api... that sucks

[crisp dew]

igpu to be corrected: 780m
is there a reason you want run Applio locally?

your setup doesn't seem powerful

cloud would be easier and faster for you

[jaunty flame]

well, because i like to experiment stuff

it's true that cloud would be much easier, but it lack the fun of it

[jaunty flame]

haizz, try to run it with the rock: fail

[crisp dew]

Maybe @runic raft could help you more with zluda, but I'm not even sure if your GPU would be good enough tbh

[jaunty flame]

well, i don't even run heavy model, it should be fine, cuz i don't train

[jaunty flame]

so he uses amd gpu huh

[crisp dew]

Even if you only do inferece, that integrated GPU is isn't good at all for any local ai tasks

[jaunty flame]

well, i tried it on stable diffusion already, it should be fine

[runic raft]

@jaunty flame there are torch builds for windows now for gfx1103

although for for gfx1103 specifically

env\python -m pip install --index-url https://rocm.nightlies.amd.com/v2/gfx110X-dgpu/ --pre torch torchaudio torchvision

if you want to use Zluda you need to use hip sdk 6.2.4 + https://github.com/likelovewant/ROCmLibs-for-gfx1103-AMD780M-APU/releases/download/v0.6.4.2/rocm.gfx1103.for.hip.6.4.2.7z

[runic raft]

@jaunty flame also when you run core.py cli you need to run it using zluda\zluda -- env\python core.py

see run-applio-amd.bat

[jaunty flame]

interesting, i using https://github.com/scottt/rocm-TheRock/releases/tag/v6.5.0rc-pytorch-gfx110x from this, but not try to use that before

[runic raft]

those are pretty old

[jaunty flame]

well, i'll try it out

[jaunty flame]

(C:\AI\Paimon\Applio\env) C:\AI\Paimon\Applio>python -m pip install --index-url https://rocm.nightlies.amd.com/v2/gfx110X-dgpu/ --pre torch torchaudio torchvision
Looking in indexes: https://rocm.nightlies.amd.com/v2/gfx110X-dgpu/
Requirement already satisfied: torch in c:\ai\paimon\applio\env\lib\site-packages (2.7.0a0+git3f903c3)
Requirement already satisfied: torchaudio in c:\ai\paimon\applio\env\lib\site-packages (2.7.0a0+52638ef)
Requirement already satisfied: torchvision in c:\ai\paimon\applio\env\lib\site-packages (0.22.0+9eb57cd)
Requirement already satisfied: filelock in c:\ai\paimon\applio\env\lib\site-packages (from torch) (3.19.1)
Requirement already satisfied: typing-extensions>=4.10.0 in c:\ai\paimon\applio\env\lib\site-packages (from torch) (4.15.0)
Requirement already satisfied: sympy>=1.13.3 in c:\ai\paimon\applio\env\lib\site-packages (from torch) (1.14.0)
Requirement already satisfied: networkx in c:\ai\paimon\applio\env\lib\site-packages (from torch) (3.5)
Requirement already satisfied: jinja2 in c:\ai\paimon\applio\env\lib\site-packages (from torch) (3.1.6)
Requirement already satisfied: fsspec in c:\ai\paimon\applio\env\lib\site-packages (from torch) (2025.9.0)
Requirement already satisfied: numpy in c:\ai\paimon\applio\env\lib\site-packages (from torchvision) (1.26.4)
Requirement already satisfied: pillow!=8.3.*,>=5.3.0 in c:\ai\paimon\applio\env\lib\site-packages (from torchvision) (11.3.0)
Requirement already satisfied: mpmath<1.4,>=1.1.0 in c:\ai\paimon\applio\env\lib\site-packages (from sympy>=1.13.3->torch) (1.3.0)
Requirement already satisfied: MarkupSafe>=2.0 in c:\ai\paimon\applio\env\lib\site-packages (from jinja2->torch) (3.0.3)
Hmm, it's the same

[runic raft]

you need to uninstall old packages

[jaunty flame]

oh right, forgot it

[jaunty flame]

wait, sdk 6.2.4? not 6.4.2?

[runic raft]

6.4.2 that one

[jaunty flame]

i tried to follow this guide but not working, i'll tried again by fresh download, do you recommend the git one or precompiled one?

[jaunty flame]

C:\AI\Paimon\Applio\env>"C:\AI\Paimon\Applio\env\python.exe" -c "import torch; print(torch.cuda.is_available()); print(torch.cuda.device_count())"
True
1
hmm

what do i miss here

C:\AI\Paimon\Applio\env>"C:\AI\Paimon\Applio\env\python.exe" -c "import torch; print('CUDA/ZLUDA Available:', torch.cuda.is_available()); print('Number of GPUs:', torch.cuda.device_count()); print('GPU Name:', torch.cuda.get_device_name(0))"
CUDA/ZLUDA Available: True
Number of GPUs: 1
GPU Name: AMD Radeon 780M Graphics
welp, it does recognize my gpu though

what is wrong here i supposed

[runic raft]

if you're using zluda you need to run it the same way run-applio-amd.bat does

set HIP_VISIBLE_DEVICES="0"
set ZLUDA_COMGR_LOG_LEVEL=1
SET DISABLE_ADDMM_CUDA_LT=1
zluda\zluda.exe -- env\python.exe

if you're using theRock wheels you should not be using zluda

to see the actual error you need to look at the console window that shows the error stack

[jaunty flame]

yeah, actually the rock wheels doesn't work when try it with applio in git, not the precompiled one

[jaunty flame]

it doesn't show anything, it just say click to exit

[runic raft]

okay, so it probably says something about miopen and not being able to find kernels

[jaunty flame]

yup, that's the one

the miopen

[runic raft]

okay, use zluda method then

at least there are libraries available

[jaunty flame]

yeah, i'm on it right now

[jaunty flame]

big first step
C:\AI\Paimon>call conda activate paimon
[INFO] Starting TTS server in the background...
[INFO] Waiting for TTS server to initialize... (This may take a minute on first run)
✅ [INFO] TTS server is running and ready.
[INFO] Starting RVC server (Zluda/ROCm) in the background...
[INFO] Waiting for RVC server to initialize models... (This is the long wait!)

✅ [INFO] RVC server is running and models are loaded.

==================================================
🚀 RVC ENVIRONMENT CHECK 🚀
Status: READY
Model Loaded: ✅ Yes
Zluda/ROCm Env: Detected
Status: RVC is running successfully under Zluda/ROCm emulation.

[jaunty flame]

fuck, when i try to run this but the pitch extraction method doesn't allow using zluda/rocm emulation, or else it has error RuntimeError: cuDNN error: CUDNN_STATUS_EXECUTION_FAILED

the pitch extraction method bring my entire code to force back to cpu

haizzzz

[jaunty flame]

i run out of option now, either run my entire code on cpu or nothing

[runic raft]

core.py extract

[jaunty flame]

What do you mean

[runic raft]

every applio functionality can be ran using core.py

or directly

env\python rvc\train\preprocess\preprocess.py X:\Applio\logs\Book G:\Training\book3 32000 4 Simple False False 0 3.0 0.3 post

env\python rvc\train\extract\extract.py X:\Applio\logs\book rmvpe 4 0 32000 spin none 2

`env\python rvc\train\train.py VCTK_32k_spin_retune 5 50 rvc\models\pretraineds\hifi-gan\f0G32k_emb129.pth rvc\models\pretraineds\hifi-gan\f0D32k.pth 0 16 32000 False True False False 5 False "HiFi-GAN" False

`

example

[jaunty flame]

Yeah, but the problem here is the pitch extraction is tied to cuDNN, and i can't run it even with zluda

I even try to seperate pitch extraction to run with cpu, the rest run on zluda but it doesn't work either

I even manually edit the pipeline.py, but it creates more problem

I can't even run core.py without pitch extraction, because it's in the core of applio

[runic raft]

that's the purpose of rvc\lib\zluda.py

torch.backends.cudnn.enabled = False

rvc\train\extract\extract.py includes that file

with latest zluda you need to replace zluda.py with a shorter patch

infer.py does not include zluda.py, so you need to add that manually

[jaunty flame]

hmm, zluda.py huh

i haven't try that

well, no use

oh wait

seem about right, because my zluda file is exactly like this when the error occur

and i need to modify the extract.py to run the pitch extraction on only cpu

intesting, i'll try it later

[jaunty flame]

modify the extract.py, to pipeline.py, what else now huh

[runic raft]

pitch extractor for training?

it already includes zluda.py

infer.py needs an import

with hip sdk 6.2.4+ you need a simple version of zluda,py from the text file above

[jaunty flame]

no, for send a text generate by llm to my rvc v2 model, and let applio send me a voice

[runic raft]

okay, so just a voice change

import zluda.py before you do from rvc.infer.infer import VoiceConverter

[jaunty flame]

well, speech to speech, i say something, it send to llm, llm answer, the answer tranfer to rvc

and it will generate a voice based on that text through tts and my rvc model

already done that, still, it's mandatory to include pitch extraction

[runic raft]

no

[jaunty flame]

something like this

[runic raft]

import rvc.lib.zluda

from rvc.infer.infer import VoiceConverter
infer_pipeline = VoiceConverter()

pth_path = r"logs\BillBurr\billbur.pth"
index_path = r""
input_path = r"X:\Applio\assets\audios\test_sample.wav"
embedder_model = "contentvec"
sid = 0

output_path = r"T:\test2.wav"
torch.manual_seed(1234)
infer_pipeline.convert_audio(audio_input_path=input_path,audio_output_path=output_path,model_path=pth_path,index_path=index_path,embedder_model=embedder_model,sid=sid)

[jaunty flame]

hmm

[runic raft]

like this

[jaunty flame]

yeah, that's voice changer for sure

[runic raft]

edge tts

import asyncio
import edge_tts

#tts settings
input_text = "tts.txt"
speaker = "en-GB-LibbyNeural"
rate = 0
input_path = "edge_out.wav"

async def main():
    rates = f"+{rate}%" if rate >= 0 else f"{rate}%"
    start_time1 = time.time()
    await edge_tts.Communicate(
        text,
        speaker,
        rate=rates,
    ).save(input_path)
    elapsed_time = time.time() - start_time1
    print(f"TTS gen time in {elapsed_time:.2f} seconds.")

if __name__ == "__main__":

    with open(input_text, 'r') as file:
        text = file.read()

    asyncio.run(main())

[jaunty flame]

but like i said, my procedure is i speak to my micro, then use stt to transfer it to text then send to llm, groq api to be exact, then the anwser of that llm will send back to tts model to became voice, then from that voice i changed it to my preference voice using my rvc model

or is there any smarter ways to do that?

[runic raft]

that's about right

[jaunty flame]

oh wait, actually i don't really need that pitch extraction

damn it, i have been so focusing on this problem that i took so long to realize that i don't need pitch extraction, i can you different method to yield similar result

[jaunty flame]

thanks for giving me an idea

[runic raft]

pitch extraction is done in the pipeline automatically

you dont need anything else

[jaunty flame]

do you think of something to replace it, or just voicechanger with my rvc model alone

[runic raft]

a tts with voice cloning, depends on the language you need

edge tts requires internet connection and everything is going to Microsoft for processing

[jaunty flame]

well, that's no brainer, i use groq api that already required internet connection already

[jaunty flame]

now, the problem is in order to connect to core.py via cli, you must connect to pipeline.py

haizzz

whether i like it or not, it always use pitch extraction by default, which my gpu doesn't like at all

[runic raft]

i gave you an example how to do VC without core.py

[jaunty flame]

you mean that voice converter?

[runic raft]

yes

[jaunty flame]

yeah, i know it, but it have pipeline.py on it, whenever my code encounter with pipeline, it crash

fuck pitch extraction

[runic raft]

it should not, as long as you've imported torch and disabled cudnn before impoting vc

[jaunty flame]

haizz, it's no use though

i changed to gradio api, and unfortunely it works, while cli doesn't

how irony

just because of one single pipeline.py

[runic raft]

what is the error? you're doing something wrong

[jaunty flame]

same old same old, the problem with cudnn, and traceback to pipeline.py