[Bug]: Major issues with transformers version causing rubbish generations with Gemma3 family using vllm

[AbdelrahmanHagrass]

Your current environment

.

🐛 Describe the bug

Major issues with transformers version when used with Gemma3 family models on vllm. The output generations are incorrect and not usable—appears to be due to incompatibility or regression in transformers. Please advise on compatible versions or fixes.

Example: Generations are rubbish or meaningless compared to expected outputs.

Tested with latest vllm and Gemma3 models.

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[DarkLight1337]

Which version of transformers are you using?

[AbdelrahmanHagrass]

The latest one and must downgrade to 4.51.1 to work fine with gemma , I don’t know why this happens but it was a major issue , we didn’t know exactly what went wrong with the newer versions

[DarkLight1337]

cc @hmellor @Isotr0py

[Isotr0py]

Which vLLM version are you using? I remember we have supported the new sliding window config introduced after Transformers v4.53 in #20541.

[AbdelrahmanHagrass]

The latest one, but not the dev , downgraded to 0.8.5.post1 , and transfomers 4.51.1 , it worked , but it took a lot of time to get compatible versions, what update happened that missed it up

[AbdelrahmanHagrass]

These guys had similar issue, but still open issue

#20341