Anyone get it running with 50xx / CUDA 12.8?

[siriShah302]

Now that we've got torch/torchaudio playing nicely together on https://download.pytorch.org/whl/cu128, has anyone figured out how to make all the dependencies play nicely together?

[quasiblob]

I too would be interested to hear if this is solvable - I tried to get xtts-webui to work with 50-series GPU, for several hours but failed. Seems like torch versions 2.6 or never will give errors when starting the app, errors about weights only loading, because of some stricter security features of never torch versions I think. Then again, I don't think there are older torch versions available for Cuda toolkit 12.8 and RTX 50-series GPUs? I'm complete Python noob so I'm mostly guessing.

[quasiblob]

I couldn't get this installed in WSL2 either - on Linux subsystem I only had to remove some package versions from pip requirement.txt to make it install all the requirements - so compared to Window install it was very easy (in Windows I also got deepspeed compiling related errors). But the app crashes when I try to run app.py.

Seems like there is some mismatch of values and what the system expects (coqpit.py)? I get value errors when trying to run the app. I think the cause for this is config file (/models/v2.0.2/config.json), and some float values are not working with the script for some reason, but I could be wrong, as I don't know anything about these things.

I tried fixing these parts of the code with the help of ChatGPT, but I couldn't get it fixed, and I don't see any point continuing.

Feels like this repository needs an update or better docs, but I guess the author isn't focusing on this GUI any longer.

[quasiblob]

⭐ Finally managed to make this work ⭐ with RTX 50 series!?

There were several issues - and I doubt that everything would work.
I have no time to test this, maybe someone else can also try this?

I used Python 3.10 (at least it worked?)

• So create your venv with 3.10, if your system has several Pythons, for me it is this:
  C:\Pythons\Python310\python.exe -m venv venv

RTX 50 series you have to install Pytorch 2.7 CUDA 12.8 version

• You can't AFAIK use older version
  pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

Very fragile requirements.txt setup

• There are conflicts (coqui, ctranslate2, coqpit, transformers)
• Lots of banging head against a wall
• I used the resulting setup to create new requirements, not the original one (see the end of this post)

Install requirements:
pip install -r requirements.txt

At least one script breaks because of Pytorch 2.7

• Because of stricter safety features of 2.7. So I had to modify \venv\lib\site-packages\trainer\io.py,

I added the "weights_only=False" part, line 83:
return torch.load(f, map_location=map_location, weights_only=False, **kwargs)

Deepspeed build fails (even if I had MS Visual Studio Build Tools 2022 version installed)

• LLMs helped, and this allowed me to build the deepspeed. Note - the latest version from github, not from pypi:

set DS_BUILD_AIO=0
set DS_BUILD_GDS=0
set DS_BUILD_OPS=0
set DS_BUILD_EVOFORMER_ATTN=0
set DS_BUILD_FP_QUANTIZER=0
pip install git+https://github.com/microsoft/DeepSpeed.git

• Note - I also installed these build tools before previous step, not sure if these were needed:
  pip install pybind11 ninja setuptools wheel

• Note - I still couldn't use --deepspeed to start the app, but at least the app now starts

"Could not locate cudnn_ops_infer64_8.dll. Please make sure it is in your library path!"

• This happens in voice2voice - seems like app requests exact version from cuDNN 8.7.6:
  cudnn_ops_infer64_8.dll

• This can be found gated behind Nvidia developer account, in Nvidia's cuDNN archive:
  Nvidia cudnn archive

• I got the files and placed them in here (not there right folder probably):
  C:\Program Files\NVIDIA\CUDNN\v8.9.7

• Note, there are other places (like Hugging Face), where you can find copies of these files, but use caution

• After this I also added its bin folder into system Path, above other Nvidia cuda entries:

C:\Program Files\NVIDIA\CUDNN\v8.9.7\bin

• After this, voice2voice no longer complained about missing file, and started working

My requirements.txt (have to double check tomorrow, as I did this twice) - created with pip freeze:

Edit - I did more experimenting with pip tools, these packages installed without any conflicts:

--extra-index-url https://download.pytorch.org/whl/cu128

absl-py==2.2.2
aiofiles==23.2.1
aiohappyeyeballs==2.6.1
aiohttp==3.11.18
aiosignal==1.3.2
alembic==1.15.2
altair==5.5.0
annotated-types==0.7.0
antlr4-python3-runtime==4.9.3
anyascii==0.3.2
anyio==4.9.0
asteroid-filterbanks==0.4.0
async-timeout==5.0.1
attrs==25.3.0
audioread==3.0.1
av==14.4.0
babel==2.17.0
blis==0.7.11
catalogue==2.0.10
certifi==2025.4.26
cffi==1.17.1
charset-normalizer==3.4.2
click==8.1.8
cloudpathlib==0.21.1
colorama==0.4.6
coloredlogs==15.0.1
colorlog==6.9.0
confection==0.1.5
contourpy==1.3.2
coqpit==0.0.17
coqpit-config==0.2.0
coqui-tts==0.24.2
coqui-tts-trainer==0.1.4
cryptography==45.0.2
ctranslate2==4.4.0
cycler==0.12.1
cymem==2.0.11
cython==3.1.0
dateparser==1.1.8
decorator==5.2.1
deepl==1.22.0
dill==0.4.0
docopt==0.6.2
einops==0.8.1
encodec==0.1.1
exceptiongroup==1.3.0
exejs==0.0.4
fastapi==0.115.12
faster-whisper==1.1.0
ffmpeg-python==0.2.0
ffmpy==0.5.0
filelock==3.18.0
flatbuffers==25.2.10
fonttools==4.58.0
frozenlist==1.6.0
fsspec[http]==2025.3.2
future==1.0.0
gradio==3.50.2
gradio-client==0.6.1
greenlet==3.2.2
grpcio==1.71.0
gruut[de,es,fr]==2.4.0
gruut-ipa==0.13.0
gruut-lang-de==2.0.1
gruut-lang-en==2.0.1
gruut-lang-es==2.0.1
gruut-lang-fr==2.0.2
h11==0.16.0
hangul-romanize==0.1.0
httpcore==1.0.9
httpx==0.28.1
huggingface-hub==0.31.2
humanfriendly==10.0
hyperpyyaml==1.2.2
idna==3.10
importlib-resources==6.5.2
inflect==7.5.0
jh2==5.0.9
jinja2==3.1.6
joblib==1.5.0
jsonlines==1.2.0
jsonschema==4.23.0
jsonschema-specifications==2025.4.1
julius==0.2.7
kiwisolver==1.4.8
langcodes==3.5.0
langid==1.1.6
language-data==1.3.0
lazy-loader==0.4
librosa==0.11.0
lightning==2.5.1.post0
lightning-utilities==0.14.3
llvmlite==0.44.0
loguru==0.7.3
lxml==5.4.0
mako==1.3.10
marisa-trie==1.2.1
markdown==3.8
markdown-it-py==3.0.0
markupsafe==2.1.5
matplotlib==3.10.3
mdurl==0.1.2
more-itertools==10.7.0
mpmath==1.3.0
msgpack==1.1.0
multidict==6.4.3
multiprocess==0.70.18
murmurhash==1.0.12
narwhals==1.39.1
networkx==3.4.2
niquests==3.14.0
nltk==3.9.1
noisereduce==3.0.3
num2words==0.5.14
numba==0.61.2
number2text==0.0.1
numpy==1.26.4
omegaconf==2.3.0
onnxruntime==1.22.0
optuna==4.3.0
orjson==3.10.18
packaging==24.2
pandas==2.2.3
pathos==0.3.4
pedalboard==0.9.16
pillow==10.4.0
platformdirs==4.3.8
pooch==1.8.2
pox==0.3.6
ppft==1.7.7
praat-parselmouth==0.4.5
preshed==3.0.9
primepy==1.3
propcache==0.3.1
protobuf==6.31.0
psutil==7.0.0
pyannote-audio==3.3.2
pyannote-core==5.0.0
pyannote-database==5.1.3
pyannote-metrics==3.2.1
pyannote-pipeline==3.0.1
pycparser==2.22
pydantic==2.11.4
pydantic-core==2.33.2
pydub==0.25.1
pygments==2.19.1
pyparsing==3.2.3
pypinyin==0.54.0
pyreadline3==3.5.4
pysbd==0.3.4
pysubs2==1.8.0
python-crfsuite==0.9.11
python-dateutil==2.9.0.post0
python-multipart==0.0.20
pytorch-lightning==2.5.1.post0
pytorch-metric-learning==2.8.1
pytz==2025.2
pyyaml==6.0.2
qh3==1.5.1
referencing==0.36.2
regex==2024.11.6
requests==2.32.3
rich==14.0.0
rpds-py==0.25.0
ruamel-yaml==0.18.10
ruamel-yaml-clib==0.2.12
safetensors==0.5.3
scikit-learn==1.6.1
scipy==1.15.3
semantic-version==2.10.0
semver==3.0.4
sentencepiece==0.2.0
shellingham==1.5.4
silero-tts==0.0.5
six==1.17.0
smart-open==7.1.0
sniffio==1.3.1
sortedcontainers==2.4.0
soundfile==0.13.1
soxr==0.5.0.post1
spacy[ja]==3.7.5
spacy-legacy==3.0.12
spacy-loggers==1.0.5
speechbrain==1.0.3
sqlalchemy==2.0.41
srsly==2.5.1
starlette==0.46.2
sudachidict-core==20250129
sudachipy==0.6.10
sympy==1.14.0
tabulate==0.9.0
tensorboard==2.19.0
tensorboard-data-server==0.7.2
tensorboardx==2.6.2.2
thinc==8.2.5
threadpoolctl==3.6.0
tokenizers==0.19.1
torch==2.7.0+cu128
torch-audiomentations==0.12.0
torch-pitch-shift==1.2.5
torchaudio==2.7.0+cu128
torchmetrics==1.7.1
torchvision==0.22.0+cu128
tqdm==4.67.1
transformers==4.42.4
translators==5.9.9
typeguard==4.4.2
typer==0.15.4
typing-extensions==4.13.2
typing-inspection==0.4.0
tzdata==2025.2
tzlocal==5.3.1
urllib3==2.4.0
urllib3-future==2.12.921
uvicorn==0.34.2
wasabi==1.1.3
wassima==1.2.2
weasel==0.4.1
websockets==11.0.3
werkzeug==3.1.3
whisperx==3.3.1
win32-setctime==1.2.0
wrapt==1.17.2
yarl==1.20.0

[djObsidian]

Thanks! Your instructions for building deepspeed worked well, i got stuck on this moment.
Added it to fork with small patches in code and updated installation process. Fixed create_multiple_reference crash (because of newer gradio probably) and proper loading of models with metadata in Pytorch 2.7.
https://github.com/djObsidian/xtts-webui

Tested text2voice from reference and multireference - works fine

[siriShah302]

@quasiblob @djObsidian Wow this is great, thanks for all the hard work sharing your progress! This is probably more of a deepspeed thing but curious if you came across it - when attempting to install deepspeed via the instructions here and on the fork, I see this:

LINK : fatal error LNK1181: cannot open input file 'aio.lib'

Curious if either of you or anyone else has come across this error on Windows 11? Again, likely not a xtts-webui issue but just wanted to get your thoughts.

[djObsidian]

Could you elablorate, at which step of deepspeed install do you get this? Try installing from this commit
deepspeedai/DeepSpeed@41fcead
There might be some breaking changes after

[zarles1]

Incredible work! I haven't tested it but will do so soon. Any chance you guys could make this work for the finetune webui? Or are there other training capabilities within your xtts-webui?

Edit:

Ok, I've tried it, and despite lots of error messages during the installation everything seems to work properly. Thank you so much! Now only the question of proper model training (via finetune-webui?) remains and I'm too dumb to transfer the steps to that webui.