Does the grammar fully enforce constraints or just increases probabilities?

[Mihaiii]

I have a prompt and a grammar that basically constraints the LLM to chose from a list of options.

Today I noticed that it doesn't work quite as it should in a sense that from time to time if outputs text that is not in the list from the grammar. I don't have clear steps to replicate (this is why I didn't opened an issue).

I don't remember this happening ~2 weeks ago (worked as expected meaning is was always reliable).

I do sampling on GPU (I mentioned this because there are multiple GPU related commits in the last 2 weeks).

Does anyone also have this issue?
Is it a regression or grammar works by just increasing probabilities without fully enforcing the constraints?

[Mihaiii]

Never mind, it was my mistake: it was one of those silly small little things you spent hours on.