Assistance Needed: Unrecognized Model Error in AutoTokenizer.from_pretrained for Gemma-2-2B

Beginners
1

Hello everyone,

I’m encountering an issue when trying to load my locally stored model (Gemma-2-2b) in my FastAPI application. Specifically, when I call:

tokenizer = AutoTokenizer.from_pretrained(
“./models/Gemma-2-2b”,
trust_remote_code=True,
use_auth_token=“MY_Token”
)
model = AutoModelForCausalLM.from_pretrained(
“./models/Gemma-2-2b”,
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
device_map=“auto” if torch.cuda.is_available() else None,
trust_remote_code=True,
use_auth_token=“MY_Token”
)

I receive the following error:

ValueError: Unrecognized model in ./models/Gemma-2-2b. Should have a model_type key in its config.json, or contain one of the following strings in its name: albert, align, altclip, aria, aria_text, audio-spectrogram-transformer…

What I’ve Tried:

I verified that my directory structure is correct.
I ensured I’m authenticated (using use_auth_token="MY_Token" and logging in via huggingface-cli login).
I attempted to patch the model’s config.json manually to add "model_type": "gemma", but that approach led to additional errors.
I also tried using trust_remote_code=True without success.

My Questions:

Does Gemma-2-2b require additional configuration beyond the standard from_pretrained() approach?
Are there known compatibility issues with the latest version of Transformers and this model?
What are the best practices for resolving model loading errors when the config.json lacks a proper model_type key?
Is there an alternative method to properly initialize this model in a FastAPI setting?

System Information:

OS: Ubuntu 22.04 LTS
Python Version: 3.10
Transformers Version: 4.48.3
Model Path: ./models/Gemma-2-2b
Framework: FastAPI + Uvicorn

Sending Love and light.

2

This is an error that often occurs when you have an old version of Transformers, but I don’t know if that’s the case.

3

Hi John,
I hope you are having a beautiful day.

Thank you for your response. Yes, you are correct, it’s certainly not the version error.

Any guesses?

Sending Love,
thelifeofjoun

4

By the way, since it seems to be used locally, I think that tokens and trust_remote_code are not necessary.

tokenizer = AutoTokenizer.from_pretrained("./models/Gemma-2-2b")
model = AutoModelForCausalLM.from_pretrained(
    "./models/Gemma-2-2b",
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    device_map=“auto” if torch.cuda.is_available() else None,
)