yes, this is indirectly hinting that during training the GPL tainted code touches every single floating point value in a model making it derivative work - even the tokenizer isn't immune to this.
A tokenizer's set of tokens isn't copyrightable in the first place, so it can't really be a derivative work of anything.
the only reason usermode is not affected is because they have an exclusion for it and only via defined communication protocol, if you go around it or attempt to put a workaround in the kernel guess what: it still violates the license - point is: it is very restrictive.
This is not how copyright law works. The GPL is a copyright license, as stated by the FSF. Something which is not subject to copyright cannot be subject to a copyright license.
The tokenizer's tokens aren't patented, for sure. They can't be trademarked (they don't identify a product or service). They aren't a trade secret (the data is public). They aren't copyrighted (not a creative work). And the GPL explicitly preserves fair use rights, so there are no contractual restrictions either.
A tokenizer is effectively a list of the top-n most common byte sequences. There's simply no basis in law for it to be subject to copyright or any other IP law in the average situation.