From 641cb595928a8914a7020e60125ed94496421801 Mon Sep 17 00:00:00 2001 From: LiJzd <3542531707@qq.com> Date: Mon, 6 Jul 2026 21:33:05 +0800 Subject: [PATCH] [Doc] Clarify fastokens availability (#45813) Signed-off-by: LjjJzd <3542531707@qq.com> Co-authored-by: OpenAI Codex --- docs/configuration/optimization.md | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/docs/configuration/optimization.md b/docs/configuration/optimization.md index c6d64b2503..0f407fbc97 100644 --- a/docs/configuration/optimization.md +++ b/docs/configuration/optimization.md @@ -276,8 +276,9 @@ By default vLLM uses the standard Hugging Face `tokenizers` library to power the fast tokenizer. For BPE tokenizers (Qwen, Llama, DeepSeek, GPT-OSS, etc.) you can switch to the [fastokens](https://github.com/crusoecloud/fastokens) Rust backend, a drop-in replacement that's substantially faster on -encode/decode and on streaming detokenization. Enable it by setting -`VLLM_USE_FASTOKENS=1`: +encode/decode and on streaming detokenization. `VLLM_USE_FASTOKENS` is +available in vLLM v0.23.0 and later. If your installed vLLM version does not +recognize the environment variable, upgrade vLLM before enabling the override: ```console VLLM_USE_FASTOKENS=1 vllm serve Qwen/Qwen3-8B