From 040c6928265062511110e1f785fcec9919e5a298 Mon Sep 17 00:00:00 2001 From: Andy Xie Date: Thu, 19 Jun 2025 00:18:11 +0800 Subject: [PATCH] [MISC] add cpu_kvcache_space_bytes to CacheConfig Signed-off-by: Andy Xie --- vllm/config.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/vllm/config.py b/vllm/config.py index 508cdfaec1c4..ce7e2a2929cf 100644 --- a/vllm/config.py +++ b/vllm/config.py @@ -1506,6 +1506,8 @@ class CacheConfig: """This enables dynamic calculation of `k_scale` and `v_scale` when kv_cache_dtype is fp8. If `False`, the scales will be loaded from the model checkpoint if available. Otherwise, the scales will default to 1.0.""" + cpu_kvcache_space_bytes: Optional[int] = None + """(CPU backend only) CPU key-value cache space.""" # Will be set after profiling. num_gpu_blocks: Optional[int] = field(default=None, init=False)