From d79855eaacfd284689fe4ff6eefc280e1bb2473c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Nicol=C3=B2=20Lucchesi?= Date: Wed, 8 Jul 2026 10:17:46 +0200 Subject: [PATCH] [Docs] `kv_sharing_fast_prefill` correction (#47044) Signed-off-by: NickLucche --- vllm/config/cache.py | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-) diff --git a/vllm/config/cache.py b/vllm/config/cache.py index aa3506af15b..1091ec5f505 100644 --- a/vllm/config/cache.py +++ b/vllm/config/cache.py @@ -161,13 +161,11 @@ class CacheConfig: """Per-DP-engine maximum concurrency at max_model_len tokens.""" kv_sharing_fast_prefill: bool = False - """This feature is work in progress and no prefill optimization takes place - with this flag enabled currently. - - In some KV sharing setups, e.g. YOCO (https://arxiv.org/abs/2405.05254), + """In some KV sharing setups, e.g. YOCO (https://arxiv.org/abs/2405.05254), some layers can skip tokens corresponding to prefill. This flag enables attention metadata for eligible layers to be overridden with metadata necessary for implementing this optimization in some models (e.g. Gemma3n) + NOTE: KV cache sharing is not supported for MRv2 (v2 model runner). """ kv_cache_memory_bytes: int | None = None