fixup local-llama service

.env

@@ -3,7 +3,7 @@ MODEL_FILE=Qwen3.6-35B-A3B-MTP-MXFP4_MOE.gguf
 N_GPU_LAYERS="99"
 
 # Доп. ключи docker-compose (опционально; дефолты заданы в compose)
-CTX_SIZE=175000
+CTX_SIZE=170000
 THREADS=12
 CACHE_TYPE_K=q8_0
 CACHE_TYPE_V=q8_0

local-llama.service

@@ -4,11 +4,11 @@ After=network.target
 
 [Service]
 Type=simple
-User=grayhook
+User=bazzite
-Group=grayhook
+Group=bazzite
-WorkingDirectory=/home/grayhook/Documents/llama-space
+WorkingDirectory=/home/bazzite/llama-space
-EnvironmentFile=/home/grayhook/Documents/llama-space/.env
+EnvironmentFile=/home/bazzite/llama-space/.env
-ExecStart=/home/grayhook/Documents/llama-space/local-llama.sh
+ExecStart=/home/bazzite/llama-space/local-llama.sh
 Restart=on-failure
 RestartSec=5
 

local-llama.sh

@@ -1,5 +1,8 @@
 #!/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
 
 /home/bazzite/Документы/llama-cpp-turboquant/build/bin/llama-server \
@@ -8,7 +11,7 @@ set -euo pipefail
   --port 8080 \
   -ngl "${N_GPU_LAYERS}" \
   --n-cpu-moe "${N_CPU_MOE}" \
-  --spec-type draft-mtp --spec-draft-n-max 3 \
+  --spec-type draft-mtp --spec-draft-n-max 2 \
   -c "${CTX_SIZE}" \
   -np 1 \
   -fa on \