`"transformers_version"` is not enforced

System Info

• transformers version: 4.21.1
• Platform: macOS-10.16-x86_64-i386-64bit
• Python version: 3.9.7
• Huggingface_hub version: 0.8.1
• PyTorch version (GPU?): 1.9.1 (False)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using GPU in script?: Yes+no
• Using distributed or parallel set-up in script?: no

Who can help?

This is a general problem with loading pretrained models in the library, not any specific model: @sgugger @stevhliu @patrickvonplaten

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, …)
• My own task or dataset (give details below)

Reproduction

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("NinedayWang/PolyCoder-2.7B")
model = AutoModelForCausalLM.from_pretrained("NinedayWang/PolyCoder-2.7B")

Expected behavior

The model’s config.json specifies that this model requires version 4.23.1 here: https://huggingface.co/NinedayWang/PolyCoder-2.7B/blob/main/config.json#L21

but if the user is trying to load this model with an older version (4.21.1), it still incorrectly and silently does load.

So, I would expect an error instead of a successful loading, because with an old version of transformers, this model produces incorrect predictions. I prefer not being able to load the checkpoint if I don’t have the required version of the library.

In other words, why do we have the "transformers_version" field in the config.json, if it is not enforced? Thanks!

[sgugger]

Ah, yet another argument for not adding config arguments like those. Thanks for pointing that out.

[urialon]

Thanks a lot!