[Docs] `kv_sharing_fast_prefill` correction (#47044)
Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai>
Nicolò Lucchesi committed
Jul 8, 2026 at 10:17 UTC
d79855eaacfd284689fe4ff6eefc280e1bb2473c
1 file changed
+2
-4
vllm/config/cache.py
+2
-4
@@ -161,13 +161,11 @@ class CacheConfig:
161
"""Per-DP-engine maximum concurrency at max_model_len tokens."""
162
163
kv_sharing_fast_prefill: bool = False
164
- """This feature is work in progress and no prefill optimization takes place
165
- with this flag enabled currently.
166
-
167
- In some KV sharing setups, e.g. YOCO (https://arxiv.org/abs/2405.05254),
164
+ """In some KV sharing setups, e.g. YOCO (https://arxiv.org/abs/2405.05254),
165
some layers can skip tokens corresponding to prefill. This flag enables
166
attention metadata for eligible layers to be overridden with metadata
167
necessary for implementing this optimization in some models (e.g. Gemma3n)
168
+ NOTE: KV cache sharing is not supported for MRv2 (v2 model runner).
169
"""
170
171
kv_cache_memory_bytes: int | None = None