[bug]: 5.10 and 5.10.1 break all the existing workarounds for the still missing RTX50 support.

[Andyx1976]

Is there an existing issue for this problem?

• I have searched the existing issues

Operating system

Windows

GPU vendor

Nvidia (CUDA)

GPU model

RTX 5080

GPU VRAM

16

Version number

5.10.1

Browser

Firefox, ui version too.

Python dependencies

Cuda Kernel Error caused by outdated pytorch.

What happened

The launcher installed version starts without any problems, except for the Cuda warning. It doesn't create any images with Cude Kernel error. That is the same for 5.8 to 5.10.1. It's thanks to the 3rd finished month of no RTX 50XX support in InvokeAI..

The old working workaround (5.8 and 5.9 versions) was entering (dev console or from a console):
uv pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128 --force-reinstall
followed sometimes (every version needed a different fix) by

uv pip install --force-reinstall --no-cache-dir bitsandbytes --no-deps

after some combination of the two, the 5.9 version started with RTX50XX working.

5.10 installs finally a recent Bits and bytes version. But any form of pytorch 128 install command (pip, pip3, uv pip) causes the launch to abort. With numPy compatibility errors. Any combination of pytorch 128 and bitsandbytes reinstalls seem to fail now. Log:

Starting up...
Started Invoke process with PID: 5284
[2025-04-20 19:16:08,809]::[InvokeAI]::INFO --> Using torch device: NVIDIA GeForce RTX 5080
[2025-04-20 19:16:21,393]::[InvokeAI]::INFO --> cuDNN version: 90701

A module that was compiled using NumPy 1.x cannot be run in
NumPy 2.1.2 as it may crash. To support both 1.x and 2.x
versions of NumPy, modules must be compiled with NumPy 2.0.
Some module may need to rebuild instead e.g. with 'pybind11>=2.12'.

If you are a user of the module, the easiest solution will be to
downgrade to 'numpy<2' or try to upgrade the affected module.
We expect that some modules will need time to support NumPy 2.

Traceback (most recent call last): File "", line 198, in run_module_as_main
File "", line 88, in run_code
File "P:\Backup\invoke.venv\Scripts\invokeai-web.exe_main.py", line 10, in
sys.exit(run_app())
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\app\run_app.py", line 62, in run_app
app, loop = get_app()
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\app\run_app.py", line 14, in get_app
from invokeai.app.api_app import app, loop
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\app\api_app.py", line 16, in
from invokeai.app.api.dependencies import ApiDependencies
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\app\api\dependencies.py", line 14, in
from invokeai.app.services.download.download_default import DownloadQueueService
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\app\services\download_init.py", line 3, in
from invokeai.app.services.download.download_base import (
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\app\services\download\download_base.py", line 13, in
from invokeai.backend.model_manager.metadata import RemoteModelFile
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\backend\model_manager_init_.py", line 9, in
from invokeai.backend.model_manager.legacy_probe import ModelProbe
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\backend\model_manager\legacy_probe.py", line 29, in
from invokeai.backend.model_manager.load.model_loaders.generic_diffusers import ConfigLoader
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\backend\model_manager\load_init_.py", line 9, in
from invokeai.backend.model_manager.load.load_base import LoadedModel, LoadedModelWithoutConfig, ModelLoaderBase
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\backend\model_manager\load\load_base.py", line 19, in
from invokeai.backend.model_manager.load.model_cache.model_cache import ModelCache
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\backend\model_manager\load\model_cache\model_cache.py", line 24, in
from invokeai.backend.model_manager.load.model_util import calc_model_size_by_data
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\backend\model_manager\load\model_util.py", line 15, in
from invokeai.backend.image_util.depth_anything.depth_anything_pipeline import DepthAnythingPipeline
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\backend\image_util_init_.py", line 12, in
from invokeai.backend.image_util.util import InitImageResizer, make_grid # noqa: F401
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\backend\image_util\util.py", line 4, in
import cv2
File "P:\Backup\invoke.venv\Lib\site-packages\cv2_init_.py", line 181, in
bootstrap()
File "P:\Backup\invoke.venv\Lib\site-packages\cv2_init_.py", line 153, in bootstrap
native_module = importlib.import_module("cv2")
File "C:\Users\AndiN1\AppData\Roaming\uv\python\cpython-3.12.9-windows-x86_64-none\Lib\importlib_init_.py", line 90, in import_module
return bootstrap.gcd_import(name[level:], package, level)
AttributeError: ARRAY_API not found
Traceback (most recent call last):
File "", line 198, in run_module_as_main
File "", line 88, in run_code
File "P:\Backup\invoke.venv\Scripts\invokeai-web.exe_main.py", line 10, in
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\app\run_app.py", line 62, in run_app
app, loop = get_app()
^^^^^^^^^
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\app\run_app.py", line 14, in get_app
from invokeai.app.api_app import app, loop
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\app\api_app.py", line 16, in
from invokeai.app.api.dependencies import ApiDependencies
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\app\api\dependencies.py", line 14, in
from invokeai.app.services.download.download_default import DownloadQueueService
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\app\services\download_init.py", line 3, in
from invokeai.app.services.download.download_base import (
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\app\services\download\download_base.py", line 13, in
from invokeai.backend.model_manager.metadata import RemoteModelFile
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\backend\model_manager_init.py", line 9, in
from invokeai.backend.model_manager.legacy_probe import ModelProbe
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\backend\model_manager\legacy_probe.py", line 29, in
from invokeai.backend.model_manager.load.model_loaders.generic_diffusers import ConfigLoader
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\backend\model_manager\load_init.py", line 9, in
from invokeai.backend.model_manager.load.load_base import LoadedModel, LoadedModelWithoutConfig, ModelLoaderBase
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\backend\model_manager\load\load_base.py", line 19, in
from invokeai.backend.model_manager.load.model_cache.model_cache import ModelCache
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\backend\model_manager\load\model_cache\model_cache.py", line 24, in
from invokeai.backend.model_manager.load.model_util import calc_model_size_by_data
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\backend\model_manager\load\model_util.py", line 15, in
from invokeai.backend.image_util.depth_anything.depth_anything_pipeline import DepthAnythingPipeline
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\backend\image_util_init.py", line 12, in
from invokeai.backend.image_util.util import InitImageResizer, make_grid # noqa: F401
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "P:\Backup\invoke.venv\Lib\site-packages\invokeai\backend\image_util\util.py", line 4, in
import cv2
File "P:\Backup\invoke.venv\Lib\site-packages\cv2_init_.py", line 181, in
bootstrap()
File "P:\Backup\invoke.venv\Lib\site-packages\cv2_init_.py", line 153, in bootstrap
native_module = importlib.import_module("cv2")
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\XXXX\AppData\Roaming\uv\python\cpython-3.12.9-windows-x86_64-none\Lib\importlib_init_.py", line 90, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ImportError: numpy.core.multiarray failed to import

What you expected to happen

I expect invokeai to launch and run...
Well what i really expected to happen was a launcher option "install nightly pytorch 128 instead", with all the needed stuff, a while ago. since the launcher is the "convenient easy to use" way to install invokeai.

How to reproduce the problem

See above.

Additional context

No response

Discord username

No response

[hipsterusername]

We’re not going to incorporate the installation of nightlies as part of a standard stable install.

You can direct your feedback to NVidia or PyTorch.

[Andyx1976]

We’re not going to incorporate the installation of nightlies as part of a standard stable install.

You can direct your feedback to NVidia or PyTorch.

The Launcher alternative options for 20xx or CPU only. It also has optional rc versions installs. It doesn't have to be in stable if it isn't. But you seem to actively sabotage ways to do it. So far it was just a (relatively) simple pytorch install away, so there isn't anything inherently incompatible (for months). And i'm not interested in the behind the scenes politics and animosities. We all know where we stand with Nvidia.
Comfyui works fine with RTX50 cards for (literally!) months, others have at least optional instructions (Fluxgym onetrainer) which are also just a simple pip pytorch thing btw. .

Also correct me if i'm wrong, (but i just downloaded the newer launcher from it), the website nowhere says "not compatible with RTX 50 cards".

--
anyway my own workaround: if someone has the problem (with the semi affordable 5060 out it will get a LOT worse btw) is rolling back to 5.9.1, which has a the same problem but can be fixed. Here is a quick how to:
get a current launcher that has a manual button next to the version when you install, click that, enter 5.9.1 , install it. Afterwards open the dev console (bottom left >_ and then click on "start dev console")
There enter:
uv pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128 --upgrade --force-reinstall
Should install some 3GB of stuff, then enter: uv pip install --force-reinstall --no-cache-dir bitsandbytes --no-deps
Then it launches (for me) on a 5080

Of course i still very much hope for a working fix for 5.10 that isn't "Go to Nvidia"

[hipsterusername]

I’m not aware of anything we’re actively doing to prevent a manual installation of the nightlies.

We’re open to contributions that would automate it, but otherwise we need to see a stable version come out that we can rely on for incorporating into the installer.

If you’d like to contribute, let us know.

[Andyx1976]

It seems using in the console
uv pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128 --no-deps --force-reinstall
does the trick for 5.10.1. (click >_ bottom left in the launcher, then "Start dev Console", than paste the command and wait til it's done)
The --no-deps is this versions special need. Bitsandbytes doesn't complain during launch. Gonna test a bit to find potential issues.

But frankly if it's that easy that can't be beyond the possibilities to add it to the end of what the launcher does in a separate option. It is literally one extra line.
It is just a slightly different one in most versions i did it to.

[psychedelicious]

We are not going to add support for 50xx series GPUs until PyTorch and its related dependencies stabilize their support for those GPUs.

The PyTorch builds you are using are nightlies. They are built from PyTorch's dev branch each night. What you get when you install a nightly today will be different from what you get when you install a nightly tomorrow. These are untested builds - in fact, you are the alpha tester. It could work fine for you and break for the next person who installs the nightly build, because they are getting a different build (despite using the same command).

In fact, the apparently broken installation you are running into is likely due to a change in PyTorch nightly builds - not Invoke v5.10. We actually loosened the requirements for our dependencies in the v5.10 release - meaning it is harder to end up with a broken build due to PyTorch/numpy/torchvision/etc dependency conflicts.

Ince PyTorch stabilizes and we can confidently provide a reproducible, working installation of Invoke for 50xx series GPUs, we'll add it to the launcher. Until then, you will need to manually install nightlies.

[mobius29er]

see #7934 for workaround. Just did this and confirms it works.