I am encountering a critical RuntimeError during the startup of my Dockerized FastAPI application. The error occurs specifically when sentence_transformers attempts to import transformers.models.auto.modeling_auto.

It seems related to an environment variable parsing issue (NoneType has no attribute split), but I cannot pinpoint which variable is causing the crash inside the transformers library.

Environment Details:

• Python: 3.11 (running in Docker)

• TensorFlow: 2.15.0

• tf-keras: 2.15.1

• Transformers: 4.39.3

• Sentence-Transformers: 2.7.0

• Keras (Standalone): Uninstalled (I verified that pip list only shows tf-keras).

The Error Traceback:

File "/app/modules/rag/reranker.py", line 8, in 
    from sentence_transformers import CrossEncoder
  File "/usr/local/lib/python3.11/site-packages/sentence_transformers/__init__.py", line 3, in 
    from .datasets import SentencesDataset, ParallelSentencesDataset
  ...
  File "/usr/local/lib/python3.11/site-packages/transformers/models/Transformer.py", line 2, in 
    from transformers import AutoModel, AutoTokenizer, AutoConfig, T5Config, MT5Config
  File "/usr/local/lib/python3.11/site-packages/transformers/utils/import_utils.py", line 1463, in __getattr__
    value = getattr(module, name)
  ...
RuntimeError: Failed to import transformers.models.auto.modeling_auto because of the following error (look up to see its traceback):
'NoneType' object has no attribute 'split'


What I have tried so far:

I suspected a Keras 2 vs Keras 3 conflict, so I uninstalled keras and kept tf-keras.

I added the following Environment Variables in docker-compose.yml and explicitly inside my Python script (os.environ):

os.environ["TRANSFORMERS_NO_TF"] = "1"
os.environ["TF_USE_LEGACY_KERAS"] = "1"
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"  # Suspected this was None, so I set it to -1

I verified that pip list shows tensorflow==2.15.0 and tf-keras==2.15.1.

Despite setting CUDA_VISIBLE_DEVICES to "-1", the error persists at the exact same import line. Does anyone know which specific environment variable transformers v4.39.3 parses with .split() that might be returning None in a Docker environment?