From f5f8e5bb797e5cc6d6a62830afe68676ade39d6d Mon Sep 17 00:00:00 2001
From: wmantly <wmantly@users.noreply.git.theta42.com>
Date: Tue, 15 Sep 2026 17:03:08 +0000
Subject: [PATCH] feat(cache): enable radix prefix sharing, chunk cache reuse,
 and 32GB disk-backed slot persistence

---
 scripts/start-server.sh | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/scripts/start-server.sh b/scripts/start-server.sh
index 410c4fe..90f4571 100755
--- a/scripts/start-server.sh
+++ b/scripts/start-server.sh
@@ -41,8 +41,12 @@ exec /opt/llama.cpp-xrip/build-nccl/bin/llama-server \
   -ngl 99 \
   -c 131072 \
   --parallel 1 \
+  --cache-prompt \
+  --cache-reuse 64 \
+  --cache-idle-slots \
+  --slot-prompt-similarity 0.10 \
+  --cache-ram 32768 \
   --slot-save-path /var/cache/llama-slots \
-  --cache-ram 16384 \
   --split-mode tensor \
   --flash-attn on \
   --batch-size 2048 \