feat(context): bump context window to 192K (196,608 tokens) with q4_0 KV cache

scripts/ollama-proxy.py

@@ -38,7 +38,7 @@ LLAMA_BASE  = "http://127.0.0.1:8080"      # llama.cpp server root
 MODEL_NAME  = "Qwen3.8-Uncensored"          # primary name
 BACKEND_MODEL = MODEL_NAME                  # what we send llama.cpp
 VERSION     = "0.5.4"                       # fake ollama version
-CTX_SIZE    = 131072
+CTX_SIZE    = 196608
 MAX_OUTPUT  = 65536
 KEEP_ALIVE  = 300
 

scripts/start-server.sh

@@ -39,7 +39,7 @@ exec /opt/llama.cpp-xrip/build-nccl/bin/llama-server \
   --reasoning-format deepseek \
   --numa isolate \
   -ngl 99 \
-  -c 131072 \
+  -c 196608 \
   --parallel 1 \
   --cache-prompt \
   --cache-idle-slots \
@@ -52,7 +52,7 @@ exec /opt/llama.cpp-xrip/build-nccl/bin/llama-server \
   --ubatch-size 1024 \
   --jinja \
   --threads 12 \
-  --cache-type-k q5_0 \
-  --cache-type-v q5_0 \
+  --cache-type-k q4_0 \
+  --cache-type-v q4_0 \
   --host 0.0.0.0 \
   --port 8080