Moved vars to file

compose/prod/docker-compose.inference.yml

@@ -12,14 +12,7 @@ services:
               count: all
               capabilities: [gpu]
     restart: unless-stopped
-    environment:
-      - INFERENCE_HTTP_HOST=0.0.0.0
-      - INFERENCE_HTTP_PORT=8001
-      - NVIDIA_VISIBLE_DEVICES=all
-      - WATCHFILES_FORCE_POLLING=true
-      - PYTHONPATH=/app
-      - HF_HOME=/root/.cache/huggingface
-      - HF_HUB_OFFLINE=1
+    env_file: ../../prod.env
     volumes:
       - ../../:/app
       - ../../models:/app/models