Error while using the Google collab : cannot import name 'DeepSpeedPlugin' from 'pytorch_lightning.plugins'

[Fqlox]

I tried both google collab (train from scratch or finetune with GPT-NEO) both have an issue while loading the module and gives : cannot import name 'DeepSpeedPlugin' from 'pytorch_lightning.plugins' (on the first cell while importing aitextgen)

I tried do uninstall and reinstall libraries as mentioned here. But it did not do the trick.

Is there a workaround ?

Thanks in advance.

[JosephTico]

Having same issue

[rgxb2807]

Having the same issue but locally (Ubuntu 22.04 )

[emvren]

Have the exact same issue on collab

[mpsparrow]

This may be a compatibility issue with the new version 2.0.0 of pytorch-lightning. Haven't fully tested it anything myself, but reverting to a previous version of pytorch-lightning like version 1.9.4 may be a workaround.

pip uninstall pytorch-lightning
pip install pytorch-lightning==1.9.4

Another possibility is to install directly from master. https://github.com/minimaxir/aitextgen/commit/78a0bfdfaff917cad6430d0d12b43f58a4b3226d looks like it comments out DeepSleepPlugin, which may solve the issue.

pip install git+https://github.com/minimaxir/aitextgen

https://github.com/minimaxir/aitextgen/pull/216 also seems to provide some bug fixes. Could try:

pip install git+https://github.com/scorixear/aitextgen

[analyticray]

I tried pip uninstall pytorch-lightning pip install pytorch-lightning==1.9.4

and also pip install git+https://github.com/minimaxir/aitextgen pip install git+https://github.com/scorixear/aitextgen

but still same error

[Fqlox]

This may be a compatibility issue with the new version 2.0.0 of pytorch-lightning. Haven't fully tested it anything myself, but reverting to a previous version of pytorch-lightning like version 1.9.4 may be a workaround.

pip uninstall pytorch-lightning
pip install pytorch-lightning==1.9.4

Another possibility is to install directly from master. 78a0bfd looks like it comments out DeepSleepPlugin, which may solve the issue.

pip install git+https://github.com/minimaxir/aitextgen

#216 also seems to provide some bug fixes. Could try:

pip install git+https://github.com/scorixear/aitextgen

Thanks for your answer I tried this on the google colab:

#!pip install -q aitextgen

!pip install git+https://github.com/scorixear/aitextgen

!pip uninstall pytorch-lightning
!pip install pytorch-lightning==1.9.4

and still got an issue (another one, but still an issue): ModuleNotFoundError: No module named 'pytorch_lightning.callbacks.progress.progress_bar'

It seems that the module aitextgen need to access a very specific version of pytorch lightning

[analyticray]

Hi. Is there any solution to this?

[JosephTico]

Maybe it's related to 7bfbddb7169f9546bb3f66145eb7bf4ef3a50e99, a rollback to an earlier PyTorch-lightning around 1.3.1

[Fqlox]

Maybe it's related to 7bfbddb, a rollback to an earlier PyTorch-lightning around 1.3.1

Tried installing 1.3.1 instead of 1.9.4 and got

ImportError: cannot import name 'get_num_classes' from 'torchmetrics.utilities.data' (/usr/local/lib/python3.9/dist-packages/torchmetrics/utilities/data.py)

[JosephTico]

Maybe it's related to 7bfbddb, a rollback to an earlier PyTorch-lightning around 1.3.1

Tried installing 1.3.1 instead of 1.9.4 and got

`ImportError: cannot import name 'get_num_classes' from 'torchmetrics.utilities.data' (/usr/local/lib/python3.9/dist-packages/torchmetrics/utilities/data.py)

`

Did you also a version prior to that commit I mentioned?

[darrinh]

On Ubuntu 22.04 doing pip install pytorch-lightning==1.7.0 got it working for me with the demo. Trying to use the training example gives:

pytorch_lightning.utilities.exceptions.MisconfigurationException: The provided lr scheduler LambdaLR doesn't follow PyTorch's LRScheduler API. You should override the LightningModule.lr_scheduler_step hook with your own logic if you are using a custom LR scheduler.

[darrinh]

I've tried various versions of the touch lib + aitext and none of them work as yet.

[analyticray]

I manage to resolve the error on the google colab by running this:

!pip install -qq pytorch-lightning==1.7.0 transformers==4.21.3 aitextgen==0.6.0

Please do let me know if it also solves your issue on colab?

[Fqlox]

I manage to resolve the error on the google colab by running this:

!pip install -qq pytorch-lightning==1.7.0 transformers==4.21.3 aitextgen==0.6.0

Please do let me know if it also solves your issue on colab?

It works fine in both colab and local (CPU) for me,

Thanks

[darrinh]

What distro are you using?

Thanks Darrin

On Fri, 31 Mar. 2023, 23:04 Fqlox, @.***> wrote:

I manage to resolve the error on the google colab by running this:

!pip install -qq pytorch-lightning==1.7.0 transformers==4.21.3 aitextgen==0.6.0

Please do let me know if it also solves your issue on colab?

It works fine in both colab and local (CPU) for me,

Thanks

— Reply to this email directly, view it on GitHub https://github.com/minimaxir/aitextgen/issues/215#issuecomment-1491825098, or unsubscribe https://github.com/notifications/unsubscribe-auth/AADOCT2TOXZJW7XK5547RL3W63B3FANCNFSM6AAAAAAWCO5FXA . You are receiving this because you commented.Message ID: @.***>

[VTSTech]

pip install -qq pytorch-lightning==1.7.0 made it work for me on Win10 x64 Py 3.11

[darrinh]

Thanks Veritas83

pip install -qq pytorch-lightning==1.7.0 made it work for me on Win10 x64 Py 3.11

Thanks, that combo doesn't work with python 3.8.10, which is default for ubuntu 22.04, so I need to try an newer version of python.

[starryendymion]

On Ubuntu 22.04 doing pip install pytorch-lightning==1.7.0 got it working for me with the demo. Trying to use the training example gives:

pytorch_lightning.utilities.exceptions.MisconfigurationException: The provided lr scheduler LambdaLR doesn't follow PyTorch's LRScheduler API. You should override the LightningModule.lr_scheduler_step hook with your own logic if you are using a custom LR scheduler.

i am getting the same error while training, found any fix ?

[anpoddubny]

Same issue with me (Python 3.10.8, Mac OS, pip install pytorch-lightning==1.7.0 ). Error with training

The provided lr scheduler LambdaLR doesn't follow PyTorch's LRScheduler API seems to be also discussed in issue #219

[ViCtOr-dev13]

try this : #!pip install -q aitextgen

!pip install git+https://github.com/scorixear/aitextgen

#!pip uninstall pytorch-lightning !pip install pytorch-lightning==1.9.4

[huyremy]

it'

try this : #!pip install -q aitextgen

!pip install git+https://github.com/scorixear/aitextgen

#!pip uninstall pytorch-lightning !pip install pytorch-lightning==1.9.4

it's work but I get this error: ModuleNotFoundError: No module named 'transformers.models.cpmant.configuration_cpmant'. Can you help me ? I think due to version of transformerr.

[atoqed-1]

I've tried the various suggestions here, and none of them work with both generation and fine tuning.

With this setup:

python 3.11 - Windows 64bit
torch == 2.0.1
pytorch-lightning == 2.0.6

You get this error for both text generation and fine tuning: ImportError: cannot import name 'DeepSpeedPlugin' from 'pytorch_lightning.plugins'

Downgrading pytorch-lightning to 1.9.4 gives this error:

ImportError: cannot import name '_TPU_AVAILABLE' from 'pytorch_lightning.utilities'

I've chased these errors and have had a case where text generation works (though I unfortunately lost that particular combination of packages), but fine tuning failed. I'm in need of fine tuning so this isn't good for my use case.

The requirements file suggests that a wide range of package versions will support the package, but it seems more apparent that a narrow version of packages should be recommended.

Do we have any insight into what exact package/python versions allows both text generation and fine tuning to work?

[UESTCSober]

try this : #!pip install -q aitextgen

!pip install git+https://github.com/scorixear/aitextgen

#!pip uninstall pytorch-lightning !pip install pytorch-lightning==1.9.4

1.9.4 really works

[francescoboc]

I have the same problem. Tried all solutions proposed, none of them work.

[WarpedWartWars]

i know this issue is over a year old, but i decided to find the package versions that cause this and similar import-time problems. and thus, i found this configuration:

---

import aitextgen tested on python 3.13.2 on windows 11 64bit with

• rust (for compiling tokenizers)

and these python packages:

pip<24.1
aitextgen==0.6.0
pytorch-lightning<1.8.0
transformers>=4.34.0,<4.40.0
torchmetrics<1.0.0

notes:

• pip<24.1 is due to

DEPRECATION: pytorch-lightning 1.7.7 has a non-standard dependency specifier torch>=1.9.*. pip 24.1 will enforce this behaviour change.

• transformers 4.40.0 requires tokenizers 1.19.x, which fails to build on python 3.13, due to

error: the configured Python interpreter version (3.13) is newer than PyO3's maximum supported version (3.12)
= help: please check if an updated version of PyO3 is available. Current version: 0.21.2

---

i haven't tested anything other than just importing the package, but at least that works

i haven't checked with scorixear's fork yet, but i will at some point