import pytorch_lightning fails with ValueError: transformers.models.auto.__spec__ is None

[nonstopfor]

🐛 Bug

A ValueError occurs whenever PyTorch Lightning is imported after Huggingface's Transformer. Importing Transformer after PyTorch does not lead to any errors, however.

To Reproduce

Tested in local anaconda environment:
import this:

from transformers import T5ForConditionalGeneration, T5Tokenizer
import pytorch_lightning

It seems that this issue doesn't exist on Google Colab.

Expected behavior

ValueError: transformers.models.auto.__spec__ is None

Environment

• CUDA:
  - GPU:
  - TITAN X (Pascal)
  - TITAN X (Pascal)
  - TITAN X (Pascal)
  - TITAN X (Pascal)
  - available: True
  - version: 10.1
• Packages:
  - numpy: 1.22.1
  - pyTorch_debug: False
  - pyTorch_version: 1.8.1+cu101
  - pytorch-lightning: 1.5.8
  - tqdm: 4.61.2
  - transformers: 4.15.0
• System:
  - OS: Linux
  - architecture:
  - 64bit
  - ELF
  - processor: x86_64
  - python: 3.9.5
  - version: Rename variables #124-Ubuntu SMP Fri Nov 10 18:29:59 UTC 2017

Additional context

[akihironitta]

It seems to happen with torchmetrics==0.7.0 which was released yesterday as reported in Lightning-AI/torchmetrics#770.

---

A quick workaround is to not use torchmetrics==0.7.0.

pip install "torchmetrics<0.7"

[nonstopfor]

Thanks! It is solved now.

[akihironitta]

Let's keep this issue open until it gets fixed :)

[BenoitDalFerro]

exact same issue exact same diagnostic with install as of yesterday 22/01/2022 20h51 GMT+1
torchmetrics 0.7.0 pyhd8ed1ab_0 conda-forge
downgrading to torchmetric<0.7.0 as a quick fix did the trick