migrate to router mode: use models.ini for multi-model support

local-llama.sh

@@ -5,25 +5,9 @@ export LD_LIBRARY_PATH=/usr/local/cuda-13.3/lib64${LD_LIBRARY_PATH:+:${LD_LIBRAR
 
 set -euo pipefail
 
+DIR="$(cd "$(dirname "$0")" && pwd)"
+
 /home/bazzite/Документы/llama-cpp-turboquant/build/bin/llama-server \
-  -m "models/${MODEL_FILE}" \
+  --models-preset "${DIR}/models.ini" \
-  --host 0.0.0.0 \
+  --models-dir "${DIR}/models" \
-  --port 8080 \
+  -np 1
-  -ngl "${N_GPU_LAYERS}" \
-  --n-cpu-moe "${N_CPU_MOE}" \
-  --spec-type draft-mtp --spec-draft-n-max 2 \
-  -c "${CTX_SIZE}" \
-  -np 1 \
-  -fa on \
-  --cache-type-k "turbo4" \
-  --cache-type-v "turbo3" \
-  --no-mmap \
-  --mlock \
-  --ctx-checkpoints 1 \
-  --cache-ram -1 \
-  --jinja \
-  --reasoning on \
-  --reasoning-budget -1 \
-  -b 2048 \
-  -ub 2048 \
-  --threads "${THREADS}"

models.ini

@@ -0,0 +1,24 @@
+[DEFAULT]
+host = 0.0.0.0
+port = 8080
+flash-attn = on
+chat-template = jinja
+reasoning = on
+reasoning-budget = -1
+cache-type-k = turbo4
+cache-type-v = turbo3
+no-mmap = true
+mlock = true
+ctx-checkpoints = true
+cache-ram = -1
+batch-size = 2048
+ubatch-size = 2048
+
+[Qwen3.6-35B]
+model = models/Qwen3.6-35B-A3B-MTP-MXFP4_MOE.gguf
+n-gpu-layers = 99
+n-cpu-moe = 29
+ctx-size = 170000
+threads = 12
+spec-type = draft-mtp
+spec-draft-n-max = 2