vLLM drops tokenizer_mode='slow' because Transformers v5 already made it a no-op
GitHub·low signal
vLLM PR #58545 (merged 2026-09-24) removes 'slow' from TokenizerMode. Transformers v5 removed slow tokenizers and AutoTokenizer.from_pretrained now ignores use_fast. Since vLLM requires transformers >= 5.10.4, 'slow' has behaved like 'hf' for some time. Configs that still pass tokenizer_mode='slow' will need editing once this ships.