Applio AMD | AI HUB | Page 1

obtuse gazelle Jul 27, 2025, 10:21 AM

#

Hello! I'm trying to install and run Applio with a Radeon RX 7800 XT. I followed the official documentation for AMD, but when I launch Applio, I get this error: UserWarning: CUDA initialization: CUDA unknown error - this may be due to an incorrectly set up environment, e.g. changing env variable CUDA_VISIBLE_DEVICES after program start. Setting the available devices to be zero. (Triggered internally at ..\c10\cuda\CUDAFunctions.cpp:108.)

magic shadow Jul 27, 2025, 12:57 PM

#

obtuse gazelle Hello! I'm trying to install and run Applio with a Radeon RX 7800 XT. I followed...

what's your adrenalin version?

obtuse gazelle Jul 27, 2025, 3:38 PM

#

magic shadow what's your adrenalin version?

the latest

#

which one should I use ?

magic shadow Jul 27, 2025, 3:43 PM

#

well, you probably messed up something in the applio guide

#

anyway, that one does not work with the latest adrenalin

#

grab the source code and unzip it into C:\Applio

#

https://github.com/IAHispano/Applio/archive/refs/tags/3.2.9.zip

#

if C:\Users\username\Miniconda3 exists, delete it

#

drop this into C:\Applio and run to install

📎 run-install.bat

#

run env\python to verify version 3.11 was installed

#

download these files into C:\Applio

#

https://github.com/scottt/rocm-TheRock/releases/download/v6.5.0rc-pytorch-gfx110x/torch-2.7.0a0+rocm_git3f903c3-cp311-cp311-win_amd64.whl
https://github.com/scottt/rocm-TheRock/releases/download/v6.5.0rc-pytorch-gfx110x/torchaudio-2.7.0a0+52638ef-cp311-cp311-win_amd64.whl

#

open command prompt cmd.exe in C:\Applio

#

run env\python -m pip install torch-2.7.0a0+rocm_git3f903c3-cp311-cp311-win_amd64.whl

#

and env\python -m pip install torchaudio-2.7.0a0+52638ef-cp311-cp311-win_amd64.whl

#

open rvc\train\train.py and add a line torch.backends.cudnn.enabled = False like this

#

benchmark should be set to False as well

#

after than use run-applio.bat

obtuse gazelle Jul 27, 2025, 4:11 PM

#

magic shadow drop this into C:\Applio and run to install

it does not create a run-applio.bat

magic shadow Jul 27, 2025, 4:17 PM

#

obtuse gazelle Jul 27, 2025, 4:18 PM

#

oh wow

#

It worked I think !!

#

My gpu is shown

magic shadow Jul 27, 2025, 4:26 PM

#

lmk if anything explodes

obtuse gazelle Jul 27, 2025, 4:26 PM

#

xd I'll, thanks a lot !

obtuse gazelle Jul 27, 2025, 5:02 PM

#

@magic shadow I started training and the gpu uses 100% idk if its normal

magic shadow Jul 27, 2025, 5:05 PM

#

yes, it is using gpu

obtuse gazelle Jul 27, 2025, 5:08 PM

#

ok thanks 🥲

magic shadow Jul 27, 2025, 5:10 PM

#

obtuse gazelle ok thanks 🥲

as long as the shared memory is not used you should get decent performance

#

dont use default high batch size

obtuse gazelle Jul 27, 2025, 5:11 PM

#

magic shadow dont use default high batch size

I'm using a 16 batch size for like a 17min database

magic shadow Jul 27, 2025, 5:11 PM

#

😕

obtuse gazelle Jul 27, 2025, 5:11 PM

#

XD

magic shadow Jul 27, 2025, 5:12 PM

#

4-8

obtuse gazelle Jul 27, 2025, 5:13 PM

#

I looked on the discord and ppl said more shouldnt be an issue

#

so idk

#

should i restart at 8?

magic shadow Jul 27, 2025, 5:14 PM

#

it 17 min is ~20 slices/min = 340 slices.... ideally you want no less than 50 steps/epoch

obtuse gazelle Jul 27, 2025, 5:15 PM

#

mh ok so what should I do ?

#

sorry m new

magic shadow Jul 27, 2025, 5:16 PM

#

try 4, or 6

#

see which results in a better sounding model

#

there's no rule set in stone what you should use

#

every dataset is different

obtuse gazelle Jul 27, 2025, 5:17 PM

#

ok I'll try 4, thanks !

obtuse gazelle Jul 28, 2025, 1:52 AM

#

magic shadow try 4, or 6

I don't really know what I did wrong but my graphs doesn't look like the documentation 🥲

magic shadow Jul 28, 2025, 1:53 AM

#

uncheck both

#

dont bother with old loss charts, use new avg_50

#

show all the charts

obtuse gazelle Jul 28, 2025, 1:55 AM

#

magic shadow Jul 28, 2025, 1:59 AM

#

not like this

#

collapse grads, expans loss_avg_50

#

like this

#

your kl chart is unusually high

obtuse gazelle Jul 28, 2025, 2:01 AM

#

obtuse gazelle Jul 28, 2025, 2:01 AM

#

magic shadow your kl chart is unusually high

idk what I've done

magic shadow Jul 28, 2025, 2:01 AM

#

click this under each chart

obtuse gazelle Jul 28, 2025, 2:02 AM

#

magic shadow Jul 28, 2025, 2:02 AM

#

anyway, I'd not expect anything good from this.. give a model from ~4.5k steps a test

#

it seems that you've resumed it with a different batch size

obtuse gazelle Jul 28, 2025, 2:03 AM

#

magic shadow it seems that you've resumed it with a different batch size

now ?

#

bcs I did not

magic shadow Jul 28, 2025, 2:04 AM

#

or stopped after a save + some steps, then resumed

#

this step back on the chart

obtuse gazelle Jul 28, 2025, 2:05 AM

#

oh yes

#

I did

#

Should I remake it ?

magic shadow Jul 28, 2025, 2:11 AM

#

what kind of voice is it in the dataset?

#

kl is supposed to do down under 1 pretty quickly

obtuse gazelle Jul 28, 2025, 2:14 AM

#

magic shadow what kind of voice is it in the dataset?

goku kid voice in french dub

magic shadow Jul 28, 2025, 2:23 AM

#

may wanna figure what's wrong with the dataset

obtuse gazelle Jul 28, 2025, 2:25 AM

#

ok I'll check that tomorrow

#

Thanks for your help !

mint bough Jul 28, 2025, 2:46 AM

#

obtuse gazelle ok I'll check that tomorrow

perhaps the dataset wasn't cleaned properly

lost storm Jul 28, 2025, 3:45 AM

#

Hi

obtuse gazelle Jul 28, 2025, 10:56 AM

#

Can I have an example of what the graphs should look like?

obtuse gazelle Jul 28, 2025, 10:56 AM

#

mint bough perhaps the dataset wasn't cleaned properly

I followed everything from the documentation so idk

#

maybe I'll try with another dataset

magic shadow Jul 28, 2025, 11:52 AM

#

obtuse gazelle Can I have an example of what the graphs should look like?

ideally these 3 going down

#

and converging at some value

#

fm chart going down or staying more or less flat

#

but usually it goes up... as long as it does not go more than 1/10k steps it is usually fine

#

example

#

as for norm g, as long as it does not shoot over 1k it is fine

obtuse gazelle Jul 28, 2025, 1:07 PM

#

okay thanks !

#

So if it's not like that, it's bcs of the dataset?

magic shadow Jul 28, 2025, 1:23 PM

#

or batch size

obtuse gazelle Jul 28, 2025, 3:23 PM

#

magic shadow or batch size

I've tried w 6 today and the same dataset and my graphs look like this 🤣

#

smoothing is at 0 btw

magic shadow Jul 28, 2025, 3:23 PM

#

something went wrong

obtuse gazelle Jul 28, 2025, 3:24 PM

#

seems like it

#

I don't really understand what I'm doing wrong

#

that confuse me

magic shadow Jul 28, 2025, 3:24 PM

#

it could be the beta torch wheel

#

it is still experimental stuff

obtuse gazelle Jul 28, 2025, 3:25 PM

#

oh :/

magic shadow Jul 28, 2025, 3:25 PM

#

unfortunately the working method with Zluda requires a driver rollback to 25.4.1 or 25.3.1

obtuse gazelle Jul 28, 2025, 3:25 PM

#

I can try that

magic shadow Jul 28, 2025, 3:27 PM

#

AMD cleanup utility to remove the current driver

#

then just follow the guide

#

https://docs.applio.org/applio/getting-started/installation#amd-gpu-support-windows

obtuse gazelle Jul 28, 2025, 3:28 PM

#

ok so I download applio again ?

magic shadow Jul 28, 2025, 3:28 PM

#

hip sdk 6.1 or 5,7

#

no, you can keep current

#

you'll deinstall torch and other stuff if you follow it

obtuse gazelle Jul 28, 2025, 3:28 PM

#

okok thanks

#

I'll try

magic shadow Jul 28, 2025, 3:30 PM

#

https://www.amd.com/en/resources/support-articles/faqs/GPU-601.html

#

https://www.amd.com/en/resources/support-articles/release-notes/RN-RAD-WIN-25-3-1.html

obtuse gazelle Jul 28, 2025, 3:45 PM

#

everything seems to work

#

I'll train a new one

obtuse gazelle Jul 29, 2025, 7:25 AM

#

?

magic shadow Jul 29, 2025, 8:27 AM

#

obtuse gazelle ?

that exploded with NaNs

#

make sure the driver did not automatically update to the latest

obtuse gazelle Jul 29, 2025, 8:33 AM

#

magic shadow make sure the driver did not automatically update to the latest

it did not

magic shadow Jul 29, 2025, 8:38 AM

#

🤷‍♂️

mint bough Jul 29, 2025, 10:06 AM

#

latest version of applio doesnt have that issue unless you messed up the code

mint bough Jul 29, 2025, 10:07 AM

#

obtuse gazelle ?

^

obtuse gazelle Jul 29, 2025, 10:09 AM

#

mint bough latest version of applio doesnt have that issue unless you messed up the code

wow idk I'm on 3.2.9

#

I'll download it again and retry

obtuse gazelle Jul 29, 2025, 11:05 AM

#

well

mint bough Jul 29, 2025, 11:10 AM

#

obtuse gazelle well

is it all NaN loss values in the terminal window?

#

that issue looks like the one in the old mangio RVC

obtuse gazelle Jul 29, 2025, 11:33 AM

#

mint bough is it all NaN loss values in the terminal window?

mint bough Jul 29, 2025, 11:35 AM

#

I have no clue as I see you're using AMD gpu joe_shrug

obtuse gazelle Jul 29, 2025, 11:37 AM

#

:(

#

I'm cursed

obtuse gazelle Jul 29, 2025, 11:38 AM

#

mint bough I have no clue as I see you're using AMD gpu <:joe_shrug:1159570387953258616>

thanks for your help tho !

magic shadow Jul 29, 2025, 12:06 PM

#

obtuse gazelle thanks for your help tho !

I've tested 6.2.4 install method on my spare pc

#

#

I know it works fine

obtuse gazelle Jul 29, 2025, 12:28 PM

#

I have 6.1.2

obtuse gazelle Jul 29, 2025, 12:30 PM

#

magic shadow I know it works fine

I’m down to try if I can

magic shadow Jul 29, 2025, 12:30 PM

#

📎 hip_sdk_624.txt

obtuse gazelle Jul 29, 2025, 12:33 PM

#

So I should go back to latest drivers ?

magic shadow Jul 29, 2025, 12:34 PM

#

yes

#

amd cleanup, reboot, install 25.6.1

#

reboot a few times

obtuse gazelle Jul 29, 2025, 12:36 PM

#

okay thanks !

obtuse gazelle Jul 29, 2025, 12:52 PM

#

magic shadow

in the patch file :
C:\Applio-3.2.9>rmdir /S /q zluda The system cannot find the file specified.

#

idk if it's fine

#

maybe rm -rf zluda ?

magic shadow Jul 29, 2025, 12:57 PM

#

if there's no zluda folder it is fine

#

as long as it does the rest it is fine

obtuse gazelle Jul 29, 2025, 12:57 PM

#

okok perfect

obtuse gazelle Jul 29, 2025, 2:02 PM

#

it seems to work

obtuse gazelle Jul 29, 2025, 2:33 PM

#

thanks you a lot !!

obtuse gazelle Jul 30, 2025, 6:41 AM

#

Should I use another Embedder Model because I want to train a French model ?

magic shadow Jul 30, 2025, 9:37 AM

#

obtuse gazelle Should I use another Embedder Model because I want to train a French model ?

default contentvec is fine

#

only as an experiment to see whether it helps or not

obtuse gazelle Jul 30, 2025, 9:53 AM

#

magic shadow 2) only as an experiment to see whether it helps or not

where can I find a good French embedder model ?

magic shadow Jul 30, 2025, 9:54 AM

#

there's new experimental spin that may improve some things

#

but you need a spin-trained pretrain as well

obtuse gazelle Jul 30, 2025, 9:59 AM

#

magic shadow but you need a spin-trained pretrain as well

I don't know what it is

magic shadow Jul 30, 2025, 10:00 AM

#

then use default one 🙂

obtuse gazelle Jul 30, 2025, 10:11 AM

#

magic shadow then use default one 🙂

I trained a first model on it. But when I try it, it's incomprehensible and there is some random breathing sounds

#

so I guessed it was because of that

obtuse gazelle Jul 30, 2025, 7:01 PM

#

obtuse gazelle I trained a first model on it. But when I try it, it's incomprehensible and ther...

any idea on how to fix this ?

magic shadow Jul 30, 2025, 7:02 PM

#

obtuse gazelle any idea on how to fix this ?

show the model_info.json

#

did you use a pretrain?

obtuse gazelle Jul 31, 2025, 4:57 AM

#

magic shadow show the model_info.json

{ "total_dataset_duration": "00:33:56", "total_seconds": 2036.0395624999999, "embedder_model": "contentvec", "speakers_id": 1 }

obtuse gazelle Jul 31, 2025, 4:58 AM

#

magic shadow did you use a pretrain?

I don't think

magic shadow Jul 31, 2025, 8:21 AM

#

obtuse gazelle I don't think

if you did uncheck the [x] Pretrained setting for some reason, that may explain your results. Or if you're trying to run inference using a custom embedder.

obtuse gazelle Jul 31, 2025, 8:24 AM

#

only custom pretrained was unchecked

#

and no I didn't use a custom embedder

obtuse gazelle Aug 4, 2025, 7:19 AM

#

hi ! sorry I was not at home so I didn’t try anything but what should I do ? Train again w the same settings ?

magic shadow Aug 4, 2025, 12:05 PM

#

can you run tensorboard an show what the chart look like for the model you trained?

obtuse gazelle Aug 6, 2025, 3:27 PM

#

magic shadow can you run tensorboard an show what the chart look like for the model you train...

magic shadow Aug 6, 2025, 3:33 PM

#

obtuse gazelle

your model exploded again

obtuse gazelle Aug 6, 2025, 6:26 PM

#

magic shadow your model exploded again

I saw that but I tried the model before the explosion and it was incomprehensible

magic shadow Aug 11, 2025, 8:29 PM

#

@obtuse gazelle lmk when you're online, I got a test to run

obtuse gazelle Aug 14, 2025, 8:36 AM

#

sorry I was on vacation

obtuse gazelle Aug 15, 2025, 12:31 PM

#

magic shadow <@854010250151723051> lmk when you're online, I got a test to run

I’m here all day 🫡

magic shadow Aug 15, 2025, 12:32 PM

#

obtuse gazelle I’m here all day 🫡

remind me, what did we manage to install last time? was it 7800xt and zluda?

obtuse gazelle Aug 15, 2025, 12:43 PM

#

magic shadow remind me, what did we manage to install last time? was it 7800xt and zluda?

latest drivers and 6.2.4

#

and yes 7800xt

magic shadow Aug 15, 2025, 12:45 PM

#

save these to Applio's folder, then run bench.bat

📎 bench.bat 📎 bench.py

obtuse gazelle Aug 15, 2025, 12:59 PM

#

magic shadow save these to Applio's folder, then run bench.bat

ok, I did it !

#

what should I do now?

magic shadow Aug 15, 2025, 1:01 PM

#

copy the output

obtuse gazelle Aug 15, 2025, 1:02 PM

#

this ?
C:\Applio-3.2.9>zluda\zluda.exe -- env\python.exe bench.py Using cuda torch.float32 Compilation is in progress. Please wait... linear : 0.0587s conv1d 192x192x1 : 0.2302s Compilation is in progress. Please wait... conv1d 192x768x3 : 0.3445s conv1d 768x768x1 : 1.1695s Compilation is in progress. Please wait... Compilation is in progress. Please wait... up_0 : 0.4844s up_1 : 0.6375s up_2 : 0.4737s up_3 : 0.4587s dn_0 : 0.2608s dn_1 : 0.2592s dn_2 : 0.2634s dn_3 : 0.2252s res1a : 0.4191s res1b : 17.1694s res1c : 17.3982s res2a : 0.9267s res2b : 17.1960s res2c : 17.7140s res3a : 1.4904s res3b : 17.5465s res3c : 17.7366s conv_post : 0.3962s torch.float16 linear : 0.0472s conv1d 192x192x1 : 0.1530s conv1d 192x768x3 : 0.2621s conv1d 768x768x1 : 0.2638s Compilation is in progress. Please wait... Compilation is in progress. Please wait... Compilation is in progress. Please wait... up_0 : 0.4987s up_1 : 0.5005s up_2 : 0.4999s up_3 : 0.4968s dn_0 : 0.2553s dn_1 : 0.2583s dn_2 : 0.2593s dn_3 : 0.2271s res1a : 0.3422s res1b : 17.2472s res1c : 17.3245s res2a : 0.3494s res2b : 17.6717s res2c : 17.2763s res3a : 0.3481s res3b : 17.7725s res3c : 17.9646s conv_post : 0.1600s torch.bfloat16

#

linear : 0.0598s conv1d 192x192x1 : 0.1472s conv1d 192x768x3 : 0.2555s conv1d 768x768x1 : 0.2573s up_0 : 0.4465s up_1 : 0.4557s up_2 : 0.4575s up_3 : 0.4576s dn_0 : 0.2578s dn_1 : 0.2560s dn_2 : 0.2563s dn_3 : 0.2247s res1a : 0.2942s res1b : 17.7791s res1c : 17.8637s res2a : 0.3042s res2b : 17.7352s res2c : 16.6792s res3a : 0.3150s res3b : 16.5893s res3c : 17.7460s conv_post : 0.1592s

magic shadow Aug 15, 2025, 1:03 PM

#

ah, run one more time so compilation times are not included

obtuse gazelle Aug 15, 2025, 1:03 PM

#

okok

magic shadow Aug 15, 2025, 1:03 PM

#

but anyway, I see the same problem

#

instead of 0.1

obtuse gazelle Aug 15, 2025, 1:09 PM

#

torch.bfloat16 linear : 0.0232s conv1d 192x192x1 : 0.1586s conv1d 192x768x3 : 0.2567s conv1d 768x768x1 : 0.2564s up_0 : 0.4270s up_1 : 0.4556s up_2 : 0.4543s up_3 : 0.4557s dn_0 : 0.2546s dn_1 : 0.2567s dn_2 : 0.2561s dn_3 : 0.2279s res1a : 0.3000s res1b : 17.4343s res1c : 17.4720s res2a : 0.3019s res2b : 17.6095s res2c : 17.9372s res3a : 0.3191s res3b : 17.7246s res3c : 17.8517s conv_post : 0.1583s

#

yeah same issue :/

obtuse gazelle Aug 15, 2025, 1:09 PM

#

magic shadow instead of 0.1

can I fix it ?

magic shadow Aug 15, 2025, 1:12 PM

#

https://drive.usercontent.google.com/download?id=1Gvg3hxNEj2Vsd2nQgwadrUEY6dYXy0H9&export=download&authuser=0

#

download it

#

unzip

#

then move all the folders into C:\program files\amd\rocm\6.2

#

overwriting when asked

obtuse gazelle Aug 15, 2025, 1:15 PM

#

ok done

magic shadow Aug 15, 2025, 1:15 PM

#

then in rvc\lib\zluda.py add 1==2 and
if 1==2 and torch.cuda.is_available()

#

err

#

hold on

obtuse gazelle Aug 15, 2025, 1:16 PM

#

oh

magic shadow Aug 15, 2025, 1:16 PM

#

just in bench.py first

obtuse gazelle Aug 15, 2025, 1:16 PM

#

okay

magic shadow Aug 15, 2025, 1:16 PM

#

run the bench

obtuse gazelle Aug 15, 2025, 1:17 PM

#

magic shadow then in rvc\lib\zluda.py add 1==2 and `if 1==2 and torch.cuda.is_available()`

i replace it here ? if torch.cuda.is_available() and torch.cuda.get_device_name().endswith("[ZLUDA]"):

magic shadow Aug 15, 2025, 1:17 PM

#

no, just add 1==2 and after if

obtuse gazelle Aug 15, 2025, 1:18 PM

#

okay

magic shadow Aug 15, 2025, 1:18 PM

#

to disable this check and run with cudnn

obtuse gazelle Aug 15, 2025, 1:20 PM

#

it's compiling again

#

Compilation is in progress. Please wait... Traceback (most recent call last): File "C:\Applio-3.2.9\bench.py", line 75, in <module> t = benchmark_op(layer.to(dtype), x.to(dtype)) File "C:\Applio-3.2.9\bench.py", line 23, in benchmark_op _ = op(x) File "C:\Applio-3.2.9\env\lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl return self._call_impl(*args, **kwargs) File "C:\Applio-3.2.9\env\lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl return forward_call(*args, **kwargs) File "C:\Applio-3.2.9\env\lib\site-packages\torch\nn\modules\conv.py", line 974, in forward return F.conv_transpose1d( RuntimeError: GET was unable to find an engine to execute this computation

magic shadow Aug 15, 2025, 1:24 PM

#

goddamit

obtuse gazelle Aug 15, 2025, 1:25 PM

#

🤣

magic shadow Aug 15, 2025, 1:25 PM

#

I've seen this before

obtuse gazelle Aug 15, 2025, 1:26 PM

#

magic shadow I've seen this before

I trust you 🫡

magic shadow Aug 15, 2025, 1:27 PM

#

I mean I could not fix it

#

well

obtuse gazelle Aug 15, 2025, 1:27 PM

#

oh

magic shadow Aug 15, 2025, 1:28 PM

#

there's this option though

#

https://github.com/lshqqytiger/TheRock/releases/tag/build0

obtuse gazelle Aug 15, 2025, 1:28 PM

#

magic shadow there's this option though

let's try it 🥲

#

so what should I do

magic shadow Aug 15, 2025, 1:29 PM

#

download the HIP SDK develop

#

it contains 6.5 folder

#

that goes to program files/amd/rocm

#

next to 6.2

#

change environment variables Path entry to point to 6.5/bin

#

download https://github.com/IAHispano/Applio/archive/refs/heads/main.zip

#

unzip to C:\

#

it makes Applio-main folder, run run-install.bat

#

while it is installing download torch and torchaudio wheels from the link above into Applio's folder

obtuse gazelle Aug 15, 2025, 1:35 PM

#

magic shadow while it is installing download torch and torchaudio wheels from the link above ...

into the main folder ?

magic shadow Aug 15, 2025, 1:35 PM

#

new applio-main

obtuse gazelle Aug 15, 2025, 1:38 PM

#

okok

#

while it's installing?

magic shadow Aug 15, 2025, 1:39 PM

#

while run-install.bat is running and doing its stuff

#

so you dont sit and stare at it for 5 minutes

obtuse gazelle Aug 15, 2025, 1:39 PM

#

okok

obtuse gazelle Aug 15, 2025, 1:40 PM

#

magic shadow while it is installing download torch and torchaudio wheels from the link above ...

done 🫡

magic shadow Aug 15, 2025, 1:41 PM

#

so run-install is done and you've downloaded the .whl files?

obtuse gazelle Aug 15, 2025, 1:41 PM

#

yes

magic shadow Aug 15, 2025, 1:41 PM

#

and unzipped hip sdk develop into rocm folder

obtuse gazelle Aug 15, 2025, 1:41 PM

#

and changed the path

magic shadow Aug 15, 2025, 1:41 PM

#

okay, open cmd in Applio-main

#

env\python -m pip install torch-, press tab to auto-complete, enter

#

env\python -m pip install torchaudio-, press tab to auto-complete, enter

#

env\python bench.py

obtuse gazelle Aug 15, 2025, 1:44 PM

#

ok so I move the bench.py in the new applio folder

magic shadow Aug 15, 2025, 1:44 PM

#

yes, copy it over

#

if it runs and does not blow up and numbers are not in 5s+ it should be good

obtuse gazelle Aug 15, 2025, 1:44 PM

#

xd

#

C:\Applio-main>env\python bench.py Traceback (most recent call last): File "C:\Applio-main\bench.py", line 1, in <module> import torch File "C:\Applio-main\env\Lib\site-packages\torch\__init__.py", line 274, in <module> _load_dll_libraries() File "C:\Applio-main\env\Lib\site-packages\torch\__init__.py", line 257, in _load_dll_libraries raise err OSError: [WinError 127] The specified procedure could not be found. Error loading "C:\Applio-main\env\Lib\site-packages\torch\lib\torch_hip.dll" or one of its dependencies.

magic shadow Aug 15, 2025, 1:46 PM

#

run set

#

show screenshot

#

that seems the path did not take

obtuse gazelle Aug 15, 2025, 1:48 PM

#

#

it did

magic shadow Aug 15, 2025, 1:48 PM

#

these red flagged files should be in 6.5/bin

obtuse gazelle Aug 15, 2025, 1:50 PM

#

that's all I got

magic shadow Aug 15, 2025, 1:50 PM

#

go to bin

#

#

they should be in Program Files/AMD/ROCm/6.5/bin

obtuse gazelle Aug 15, 2025, 1:51 PM

#

they are mb

magic shadow Aug 15, 2025, 1:53 PM

#

could you run where hipconfig

obtuse gazelle Aug 15, 2025, 1:53 PM

#

in the bin?

magic shadow Aug 15, 2025, 1:53 PM

#

no, from a cmd window

obtuse gazelle Aug 15, 2025, 1:53 PM

#

okok

magic shadow Aug 15, 2025, 1:54 PM

#

a new cmd window in applio-main

obtuse gazelle Aug 15, 2025, 1:54 PM

#

C:\Program Files\AMD\ROCm\6.5\bin\hipconfig.bat C:\Program Files\AMD\ROCm\6.5\bin\hipconfig.exe

magic shadow Aug 15, 2025, 1:54 PM

#

okay, now env\python bench.py

#

(assuming you did open it in applio-main)

obtuse gazelle Aug 15, 2025, 1:55 PM

#

xd no I did it in a new cmd

#

C:\Applio-main>env\python bench.py Traceback (most recent call last): File "C:\Applio-main\bench.py", line 1, in <module> import torch File "C:\Applio-main\env\Lib\site-packages\torch\__init__.py", line 274, in <module> _load_dll_libraries() File "C:\Applio-main\env\Lib\site-packages\torch\__init__.py", line 257, in _load_dll_libraries raise err OSError: [WinError 127] The specified procedure could not be found. Error loading "C:\Applio-main\env\Lib\site-packages\torch\lib\torch_hip.dll" or one of its dependencies.

#

same error

magic shadow Aug 15, 2025, 1:55 PM

#

well, open a new cmd in applio-main

obtuse gazelle Aug 15, 2025, 1:55 PM

#

okok

#

same

magic shadow Aug 15, 2025, 1:56 PM

#

oh boy

#

big guns now

obtuse gazelle Aug 15, 2025, 1:56 PM

#

🥺

magic shadow Aug 15, 2025, 1:56 PM

#

https://github.com/lucasg/Dependencies/releases/download/v1.11.1/Dependencies_x64_Release.zip

#

download anywhere, unzip, run dependencies gui

obtuse gazelle Aug 15, 2025, 1:56 PM

#

done

magic shadow Aug 15, 2025, 1:57 PM

#

file, open, C:\Applio-main\env\Lib\site-packages\torch\lib\torch_hip.dll

#

screenshot

obtuse gazelle Aug 15, 2025, 1:58 PM

#

#

all are incorrect in 6.5

magic shadow Aug 15, 2025, 1:59 PM

#

no, the top left corner

obtuse gazelle Aug 15, 2025, 1:59 PM

#

oh

magic shadow Aug 15, 2025, 1:59 PM

#

#

like this

obtuse gazelle Aug 15, 2025, 1:59 PM

#

magic shadow Aug 15, 2025, 1:59 PM

#

at the bottom, click the left column header

#

are there any like this?

obtuse gazelle Aug 15, 2025, 2:00 PM

#

no

#

i have 0 questionmark

magic shadow Aug 15, 2025, 2:12 PM

#

weirder and weirder

obtuse gazelle Aug 15, 2025, 2:12 PM

#

this is crazy 🤣

magic shadow Aug 15, 2025, 2:12 PM

#

try expanding those files at the top

#

click >

#

#

see if anything red shows under them

#

you can also try opening C:\Applio-main\env\Lib\site-packages\torch\__init__.py in notepad, find 0x0000 and change it to 0x0001

#

prev_error_mode = kernel32.SetErrorMode(0x0001)

#

oops, change it to 0x0000

#

save, then run the bench again

#

it should pop up a window to say what it could not find

obtuse gazelle Aug 15, 2025, 2:15 PM

#

magic shadow see if anything red shows under them

too much files to expand

magic shadow Aug 15, 2025, 2:16 PM

#

i did not say expand all files, just 1st level

obtuse gazelle Aug 15, 2025, 2:16 PM

#

oh nothing on 1st level

#

no red

magic shadow Aug 15, 2025, 2:16 PM

#

okay, lets try the init edit

#

change prev_error_mode = kernel32.SetErrorMode(0x0001) to prev_error_mode = kernel32.SetErrorMode(0x0000)

obtuse gazelle Aug 15, 2025, 2:18 PM

#

magic shadow Aug 15, 2025, 2:20 PM

#

okay

#

could you try setting HIP_PATH env variable

#

and point it to program files/amd/rocm/6.5

#

(not bin)

#

open new cmd after that and run the bench again

obtuse gazelle Aug 15, 2025, 2:25 PM

#

its working

#

finally

magic shadow Aug 15, 2025, 2:26 PM

#

how are the numbers?

obtuse gazelle Aug 15, 2025, 2:26 PM

#

xd

#

res1a : 0.2208s MIOpen(HIP): Warning [IsEnoughWorkspace] [GetSolutionsFallback WTI] Solver <GemmFwdRest>, workspace required: 1228800, provided ptr: 0000000000000000 size: 0 MIOpen(HIP): Warning [IsEnoughWorkspace] [EvaluateInvokers] Solver <GemmFwdRest>, workspace required: 1228800, provided ptr: 0000000000000000 size: 0 res1b : 9.5653s MIOpen(HIP): Warning [IsEnoughWorkspace] [GetSolutionsFallback WTI] Solver <GemmFwdRest>, workspace required: 1228800, provided ptr: 0000000000000000 size: 0 MIOpen(HIP): Warning [IsEnoughWorkspace] [EvaluateInvokers] Solver <GemmFwdRest>, workspace required: 1228800, provided ptr: 0000000000000000 size: 0 res1c : 9.5043s res2a : 0.5534s MIOpen(HIP): Warning [IsEnoughWorkspace] [GetSolutionsFallback WTI] Solver <GemmFwdRest>, workspace required: 2867200, provided ptr: 0000000000000000 size: 0 MIOpen(HIP): Warning [IsEnoughWorkspace] [EvaluateInvokers] Solver <GemmFwdRest>, workspace required: 2867200, provided ptr: 0000000000000000 size: 0 res2b : 21.4636s MIOpen(HIP): Warning [IsEnoughWorkspace] [GetSolutionsFallback WTI] Solver <GemmFwdRest>, workspace required: 2867200, provided ptr: 0000000000000000 size: 0 MIOpen(HIP): Warning [IsEnoughWorkspace] [EvaluateInvokers] Solver <GemmFwdRest>, workspace required: 2867200, provided ptr: 0000000000000000 size: 0

magic shadow Aug 15, 2025, 2:27 PM

#

could you do env\python bench.py > 7800xt.txt

#

and upload the file

obtuse gazelle Aug 15, 2025, 2:36 PM

#

📎 7800xt.txt

#

and in the cmd :

📎 message.txt

magic shadow Aug 15, 2025, 2:37 PM

#

omfg

#

it is even worse

obtuse gazelle Aug 15, 2025, 2:37 PM

#

xddd

magic shadow Aug 15, 2025, 2:37 PM

#

go back to zluda lol

#

change the env variables back to 6.2

#

and delete 6.5, that's 15GB of junk

obtuse gazelle Aug 15, 2025, 2:39 PM

#

😃

#

done

obtuse gazelle Aug 15, 2025, 3:46 PM

#

magic shadow go back to zluda lol

so I can't do anything? :(

magic shadow Aug 15, 2025, 3:46 PM

#

not yet

#

they need to fix this MIOpen thing

obtuse gazelle Aug 15, 2025, 3:49 PM

#

oh 🙁

#

i'll wait

magic shadow Aug 15, 2025, 3:50 PM

#

other option is to try the nightly zluda build

#

but again it is another experimental thing

obtuse gazelle Aug 15, 2025, 3:50 PM

#

xd

#

we can try

#

I like trying

magic shadow Aug 15, 2025, 3:51 PM

#

HIP SDK extension: https://drive.google.com/file/d/1Gvg3hxNEj2Vsd2nQgwadrUEY6dYXy0H9/view?usp=sharing

Google Docs

HIP-SDK-extension.zip

#

this upzips into rocm/6.2 folder

#

it should overwrite some stuff

#

you know what

#

lets not do that

obtuse gazelle Aug 15, 2025, 3:52 PM

#

oh

magic shadow Aug 15, 2025, 3:52 PM

#

i wanna do something better with my friday afternoon

obtuse gazelle Aug 15, 2025, 3:52 PM

#

xdd

#

np !

magic shadow Aug 15, 2025, 3:52 PM

#

https://github.com/lshqqytiger/ZLUDA/releases/download/rel.5e717459179dc272b7d7d23391f0fad66c7459cf/ZLUDA-nightly-windows-rocm6-amd64.zip

#

you'll need to re-do the 6.2.4 install, but use thi file instead

#

📎 hip_sdk_624.txt

#

edit patch zluda and put this file instead of the regular zluda rocm6

#

if you dont want to runin your current install, unzip the compiled version of applio somewhere else again

obtuse gazelle Aug 15, 2025, 4:03 PM

#

magic shadow https://github.com/lshqqytiger/ZLUDA/releases/download/rel.5e717459179dc272b7d7d...

I didn't really understand what I have to do w that

#

but answer when u can dw !! I don't want to ruin ur afternoon

magic shadow Aug 15, 2025, 4:04 PM

#

patch zluda 62.bat has curl -s -L https://github.com/lshqqytiger/ZLUDA/releases/download/rel.5e717459179dc272b7d7d23391f0fad66c7459cf/ZLUDA-windows-rocm6-amd64.zip > zluda.zip

#

replace it with the nightly

obtuse gazelle Aug 15, 2025, 4:06 PM

#

oh so i change the link w the nightly in the .bat

#

Should i test w bench.bat?

magic shadow Aug 15, 2025, 4:20 PM

#

did you uninstall/reinstall torch cu118?

obtuse gazelle Aug 15, 2025, 4:21 PM

#

I did

magic shadow Aug 15, 2025, 4:21 PM

#

then patched it using updated night zluda?

obtuse gazelle Aug 15, 2025, 4:21 PM

#

yes

magic shadow Aug 15, 2025, 4:22 PM

#

and hip extensions unzipped to the rocm/6.2 folder with overwriting some of the stuff there?

obtuse gazelle Aug 15, 2025, 4:22 PM

#

yes

magic shadow Aug 15, 2025, 4:22 PM

#

okay, you can try running the bench

#

one with 1==2 and one without

#

to see if that makes a difference

obtuse gazelle Aug 15, 2025, 4:26 PM

#

zluda was extracted in main and not in a folder

#

C:\Applio-3.2.9>zluda\zluda.exe -- env\python.exe bench.py The system cannot find the path specified.

#

so it says that

magic shadow Aug 15, 2025, 4:27 PM

#

ah.. no, windows defender does not like nightly

#

or firefox

obtuse gazelle Aug 15, 2025, 4:28 PM

#

defender I think

magic shadow Aug 15, 2025, 4:28 PM

#

yes, for some reason it does not have zluda folder in it

#

if you ran the patch already, you need to re-do torch again

#

since an important dll was moved

obtuse gazelle Aug 15, 2025, 4:28 PM

#

okok

magic shadow Aug 15, 2025, 4:29 PM

#

okay, so make sure the content of nightly is manually unzipped into zluda folder

#

in patch zluda delete lines

curl -s -L https://github.com/lshqqytiger/ZLUDA/releases/download/rel.5e717459179dc272b7d7d23391f0fad66c7459cf/ZLUDA-windows-rocm6-amd64.zip > zluda.zip
tar -xf zluda.zip
del zluda.zip

obtuse gazelle Aug 15, 2025, 4:40 PM

#

well well well

#

torch.float16 linear : 0.0290s conv1d 192x192x1 : 0.2268s conv1d 192x768x3 : 0.2659s conv1d 768x768x1 : 0.2734s up_0 : 0.5102s up_1 : 0.5176s up_2 : 0.5192s up_3 : 0.5293s dn_0 : 0.2644s dn_1 : 0.2683s dn_2 : 0.2643s dn_3 : 0.2320s res1a : 0.3522s res1b : 17.8821s res1c : 17.8233s res2a : 0.3527s res2b : 18.0610s res2c : 18.2808s res3a : 0.3554s res3b : 18.1599s res3c : 18.2609s conv_post : 0.1569s

magic shadow Aug 15, 2025, 4:41 PM

#

this is with 1==2?

obtuse gazelle Aug 15, 2025, 4:41 PM

#

without

magic shadow Aug 15, 2025, 4:41 PM

#

that's cudnn disabled

#

so as expected it uses a slow method

obtuse gazelle Aug 15, 2025, 4:42 PM

#

okok

#

🙁

magic shadow Aug 15, 2025, 4:51 PM

#

so?

obtuse gazelle Aug 15, 2025, 4:52 PM

#

torch.float16 linear : 0.0693s conv1d 192x192x1 : 0.1666s conv1d 192x768x3 : 0.2663s conv1d 768x768x1 : 0.2680s up_0 : 0.5272s up_1 : 0.5275s up_2 : 0.5016s up_3 : 0.5246s dn_0 : 0.2676s dn_1 : 0.2680s dn_2 : 0.2690s dn_3 : 0.2374s res1a : 0.3558s res1b : 17.8415s res1c : 18.1317s res2a : 0.3602s res2b : 18.1202s res2c : 18.1842s res3a : 0.3532s res3b : 18.2191s res3c : 18.3551s conv_post : 0.1692s

magic shadow Aug 15, 2025, 4:54 PM

#

ffs

obtuse gazelle Aug 15, 2025, 4:55 PM

#

😕

#

at least we tried

obtuse gazelle Sep 3, 2025, 8:06 PM

#

@magic shadow nothing new?

magic shadow Sep 3, 2025, 8:10 PM

#

nope, also they have some issue with nightly wheels

obtuse gazelle Sep 3, 2025, 8:11 PM

#

wow

#

I'll wait

magic shadow Sep 3, 2025, 8:15 PM

#

latest wheels are torch-2.9.0a0+rocm7.0.0rc20250826-cp311-cp311-win_amd64.whl torchaudio-2.8.0a0+rocm7.0.0rc20250826-cp311-cp311-win_amd64.whl

#

from a week ago

magic shadow Sep 10, 2025, 2:57 AM

#

there are newer wheels posted to https://rocm.nightlies.amd.com/v2/gfx110X-dgpu/

obtuse gazelle Sep 11, 2025, 8:14 PM

#

oh nice

#

what do I need to do ?

magic shadow Sep 11, 2025, 8:31 PM

#

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

#

from applio folder

obtuse gazelle Sep 11, 2025, 8:33 PM

#

from a fresh applio ?

magic shadow Sep 11, 2025, 9:50 PM

#

yeah, grab v3.4.0 while you're at it

#

@obtuse gazelle

obtuse gazelle Sep 11, 2025, 9:58 PM

#

magic shadow `env\python -m pip install torch torchaudio --index-url https://rocm.nightlies.a...

--index-url option requires 1 argument

magic shadow Sep 11, 2025, 10:08 PM

#

works for me

#

if you get 'requirement already satisfied', try env\python -m pip uninstall torch torchaudio torchvision first

obtuse gazelle Sep 12, 2025, 4:45 AM

#

magic shadow works for me

ok done

#

should I run bench.py to test ?

magic shadow Sep 12, 2025, 9:06 AM

#

obtuse gazelle should I run bench.py to test ?

you can give it a try

#

but i would not expect much.. they have not updated my ticked with any resolution

#

Applio should work without (or with very minimal changes), they are still fixing other things

obtuse gazelle Sep 12, 2025, 10:01 AM

#

magic shadow you can give it a try

Failed to load amdhip64.dll: amdhip64.dll: Can't open: The specified module could not be found. (0x7E) [ERROR] amdgpu-arch failed with return code 1 [stderr]

magic shadow Sep 12, 2025, 10:02 AM

#

bigger error

obtuse gazelle Sep 12, 2025, 10:06 AM

#

in the 3.4.0 file, there is no more amd installation?

magic shadow Sep 12, 2025, 10:07 AM

#

??

#

you download the compiled v3.4.0, unzip, uninstall torch torchaudio torchvision

#

install torch from nightly

#

that should be it

#

after that you can run the bench

#

env\python bench.py

obtuse gazelle Sep 12, 2025, 10:13 AM

#

y

#

same error when I try to run applio

magic shadow Sep 12, 2025, 10:16 AM

#

show the full error

#

you may be missing this file in C:\windows\system32

📎 libomp140.x86_64.dll

#

have you used dependency checker before?

obtuse gazelle Sep 12, 2025, 10:28 AM

#

magic shadow have you used dependency checker before?

I think so

obtuse gazelle Sep 12, 2025, 10:29 AM

#

magic shadow you may be missing this file in C:\windows\system32

so I add it ?

magic shadow Sep 12, 2025, 10:29 AM

#

you can check this file

#

but most likelt it will be the .dll I gave you

#

just save it either in the same folder as shm.dll or into windows/system32

obtuse gazelle Sep 12, 2025, 10:30 AM

#

ok it worked

#

o7

magic shadow Sep 12, 2025, 10:31 AM

#

also

#

check windows/system32 folder

#

there should be a file name amdhip64_number.dll

#

6 or 7

#

make a copy of it without _number

#

that should fix

#

as I see from the bench screenshot they have not fixed the res b/c yet

#

9 seconds is 10x slower than expected

#

but anyway, give training a try

#

may need to fix a few things manually

obtuse gazelle Sep 12, 2025, 10:34 AM

#

magic shadow make a copy of it without _number

in the same folder as shm?

magic shadow Sep 12, 2025, 10:35 AM

#

c:\windows\system32

obtuse gazelle Sep 12, 2025, 10:35 AM

#

so I copy it in system 32?

magic shadow Sep 12, 2025, 10:35 AM

#

if training explodes, delete ``` dist.init_process_group(
backend="gloo" if sys.platform == "win32" or device.type != "cuda" else "nccl",
init_method="env://",
world_size=n_gpus if device.type == "cuda" else 1,
rank=rank if device.type == "cuda" else 0,
)

magic shadow Sep 12, 2025, 10:35 AM

#

obtuse gazelle so I copy it in system 32?

yes

#

and the other possible fix needed is removing @torch.jit.script from rvc\lib\algorithm\commons.py

#

(they should fix this soon enough)

obtuse gazelle Sep 12, 2025, 10:37 AM

#

magic shadow yes

did it

#

it worked

obtuse gazelle Sep 12, 2025, 10:38 AM

#

magic shadow if training explodes, delete ``` dist.init_process_group( backend="gl...

okok I'll try a training later and I'll tell u

obtuse gazelle Sep 12, 2025, 11:49 AM

#

@magic shadow When I start training :
AttributeError: module 'torch.distributed' has no attribute 'init_process_group'

magic shadow Sep 12, 2025, 1:59 PM

#

obtuse gazelle <@155030383648440320> When I start training : `AttributeError: module 'torch.dis...

#1398973453918535780 message

obtuse gazelle Sep 12, 2025, 6:56 PM

#

magic shadow if training explodes, delete ``` dist.init_process_group( backend="gl...

magic shadow Sep 12, 2025, 7:04 PM

#

obtuse gazelle

rvc/train/train.py

#

obtuse gazelle Sep 12, 2025, 7:06 PM

#

magic shadow rvc/train/train.py

mb

#

ok done

#

I think it works, I'll try tomorrow

magic shadow Sep 17, 2025, 8:58 PM

#

devs have reproduced the slowness with res b/c tests, so hopefully going to make some workaround

#

the dist.init_process_group should've been fixed partially(?) with latest

#

https://rocm.nightlies.amd.com/v2/gfx110X-dgpu/torch-2.10.0a0%2Brocm7.0.0rc20250913-cp311-cp311-win_amd64.whl

obtuse gazelle Sep 22, 2025, 2:50 PM

#

magic shadow devs have reproduced the slowness with res b/c tests, so hopefully going to make...

I hope so

magic shadow Sep 26, 2025, 8:20 PM

#

could you try downloading the latest applio, then installing the latest nightly wheels again and try inference?

#

or training with bf16 precision

#

@obtuse gazelle

obtuse gazelle Sep 26, 2025, 8:42 PM

#

where do I find the latest nightly wheels ?

magic shadow Sep 26, 2025, 9:01 PM

#

use the same method as above

#

#1398973453918535780 message

obtuse gazelle Sep 29, 2025, 9:42 AM

#

res1b : 9.5445s MIOpen(HIP): Warning [IsEnoughWorkspace] [GetSolutionsFallback WTI] Solver <GemmFwdRest>, workspace required: 1228800, provided ptr: 0000000000000000 size: 0 MIOpen(HIP): Warning [IsEnoughWorkspace] [EvaluateInvokers] Solver <GemmFwdRest>, workspace required: 1228800, provided ptr: 0000000000000000 size: 0 res1c : 9.4794s res2a : 0.6199s MIOpen(HIP): Warning [IsEnoughWorkspace] [GetSolutionsFallback WTI] Solver <GemmFwdRest>, workspace required: 2867200, provided ptr: 0000000000000000 size: 0 MIOpen(HIP): Warning [IsEnoughWorkspace] [EvaluateInvokers] Solver <GemmFwdRest>, workspace required: 2867200, provided ptr: 0000000000000000 size: 0 res2b : 21.3799s MIOpen(HIP): Warning [IsEnoughWorkspace] [GetSolutionsFallback WTI] Solver <GemmFwdRest>, workspace required: 2867200, provided ptr: 0000000000000000 size: 0 MIOpen(HIP): Warning [IsEnoughWorkspace] [EvaluateInvokers] Solver <GemmFwdRest>, workspace required: 2867200, provided ptr: 0000000000000000 size: 0 res2c : 21.3607s

magic shadow Sep 29, 2025, 12:03 PM

#

that has not been fixed yet, I meant other things

#

bfloat16 precistion training

obtuse gazelle Dec 9, 2025, 6:23 PM

#

@magic shadow still no fix ?

magic shadow Dec 9, 2025, 6:24 PM

#

obtuse gazelle <@155030383648440320> still no fix ?

there are newer wheels

obtuse gazelle Dec 9, 2025, 6:25 PM

#

should I try?

magic shadow Dec 9, 2025, 6:26 PM

#

you can give it a try and report back

obtuse gazelle Dec 9, 2025, 6:26 PM

#

ok thanks, I'll try tomorrow

magic shadow Dec 24, 2025, 3:04 AM

#

pip install --index-url https://rocm.nightlies.amd.com/v2/gfx120X-all/ --pre torch==2.9.1+rocm7.11.0a20251223

#

you can test it in a separate virtual environment

magic shadow Dec 24, 2025, 3:04 AM

#

obtuse gazelle ok thanks, I'll try tomorrow

.

obtuse gazelle Dec 24, 2025, 3:32 PM

#

mb

#

I’m not at home, I’ll try saturday !

#Applio AMD