No Timestamp tokens emitted on inference with multilingual model?

[vade]

Hi friends. Im porting Whisper to CoreML, and attempting to implement/port the existing timestamp logic.

I'm using the multilingual 'base' model to test with. Because of this, as I understand it, I always need to initialize a sot sequence with the sot token, the lang token and the task token, as per : this line in tokenizer.py

My reading of the code implies I should be seeing timestamp tokens emitted by the model with values above special tokens, specifically "<|notimestamps|>" + 1 as per this line in tokenizer.py

I'm not seeing those tokens emitted, but I do get a no time stamp token (5036) emitted just before eot.

Do I need an additional sot sequence token to emit timestamp tokens? Ive tried appending the timestamp begin token to no avail. Or am I misapprehending and only the en models emit time stamps?

Simple question, just trying to ensure my reading is correct.

Much obliged.

[vade]

PEBKAC lol :D