fix(memory): tune cache-ram to 12288MB to prevent host RAM OOM-kills on 48GB servers

2026-09-18 02:07:18 +00:00
parent aebc82e403
commit ff29e72442
1 changed files with 1 additions and 2 deletions
@@ -42,10 +42,9 @@ exec /opt/llama.cpp-xrip/build-nccl/bin/llama-server \
  -c 131072 \
  --parallel 1 \
  --cache-prompt \
  --cache-reuse 64 \
  --cache-idle-slots \
  --slot-prompt-similarity 0.10 \
-  --cache-ram 32768 \
+  --cache-ram 12288 \
  --slot-save-path /var/cache/llama-slots \
  --split-mode tensor \
  --flash-attn on \