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

.env

@@ -1,10 +1,10 @@
-MODEL_FILE=Qwen3.6-35B-A3B-MXFP4_MOE.gguf
+MODEL_FILE=Qwen3.6-35B-A3B-MTP-MXFP4_MOE.gguf
 # Для -1 в .env лучше кавычки — иначе часть парсеров .env ломается на ведущем минусе.
-N_GPU_LAYERS="-1"
+N_GPU_LAYERS="99"
 
 # Доп. ключи docker-compose (опционально; дефолты заданы в compose)
-CTX_SIZE=65536
-THREADS=8
+CTX_SIZE=170000
+THREADS=12
 CACHE_TYPE_K=q8_0
 CACHE_TYPE_V=q8_0
 

local-llama.service

@@ -0,0 +1,25 @@
+[Unit]
+Description=llama-server (local-llama)
+After=network.target
+
+[Service]
+Type=simple
+User=bazzite
+Group=bazzite
+WorkingDirectory=/home/bazzite/llama-space
+EnvironmentFile=/home/bazzite/llama-space/.env
+ExecStart=/home/bazzite/llama-space/local-llama.sh
+Restart=on-failure
+RestartSec=5
+
+# CPU pinning: taskset -c 0-11
+CPUAffinity=0-11
+
+# ulimit -l unlimited → lock memory (mlockall / MADV_POPULATE_READ + mlock)
+LimitMEMLOCK=infinity
+
+# Не убивать OOM killer — модель критична
+OOMScoreAdjust=-900
+
+[Install]
+WantedBy=multi-user.target

local-llama.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+
+export PATH=/usr/local/cuda-13.3/bin${PATH:+:${PATH}}
+export LD_LIBRARY_PATH=/usr/local/cuda-13.3/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
+
+set -euo pipefail
+
+DIR="$(cd "$(dirname "$0")" && pwd)"
+
+/home/bazzite/Документы/llama-cpp-turboquant/build/bin/llama-server \
+  --models-preset "${DIR}/models.ini" \
+  --models-dir "${DIR}/models" \
+  -np 1

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

podman-llama.sh

@@ -30,7 +30,7 @@ fi
 : "${CACHE_TYPE_K:=q8_0}"
 : "${CACHE_TYPE_V:=turbo2}"
 : "${THREADS:=8}"
-: "${IMAGE:=ghcr.io/ggml-org/llama.cpp:server-cuda}"
+: "${IMAGE:=ghcr.io/ggml-org/llama.cpp:server-cuda13}"
 : "${BIND_HOST:=0.0.0.0}"
 
 # Разбить по пробелам для podman (пусто = только CDI-устройство).
@@ -42,6 +42,7 @@ podman run -d \
   --replace \
   --restart no \
   --shm-size 1g \
+  --cpuset-cpus="0-11" \
   --ulimit memlock=-1:-1 \
   -p "${BIND_HOST}:${PORT}:8080" \
   -v "${ROOT}/models:/models:ro" \
@@ -52,7 +53,9 @@ podman run -d \
   -m "/models/${MODEL_FILE}" \
   --host 0.0.0.0 \
   --port 8080 \
+  -ngl "${N_GPU_LAYERS}" \
   --n-cpu-moe "${N_CPU_MOE}" \
+  --spec-type draft-mtp --spec-draft-n-max 3 \
   -c "${CTX_SIZE}" \
   -np 1 \
   -fa on \
@@ -61,7 +64,7 @@ podman run -d \
   --no-mmap \
   --mlock \
   --ctx-checkpoints 1 \
-  --cache-ram 0 \
+  --cache-ram -1 \
   --jinja \
   --reasoning on \
   --reasoning-budget -1 \