COMPOSE_PROFILES=embed LLAMACPP_REF=b6000 LLAMACPP_COMMIT=4762ad7316dcdec20016ab5985fb46a27902204d CUDA_VERSION=12.4.0 CUDA_DOCKER_ARCH=52;61 MODELS_DIR=/srv/models EMBED_GGUF=Qwen3-Embedding-0.6B-Q8_0.gguf EMBED_POOLING=last RERANK_GGUF=bge-reranker-v2-m3-Q8_0.gguf CHAT_GGUF=Qwen3-4B-Q4_K_M.gguf # Karte je Dienst: Index oder UUID aus `nvidia-smi -L` (Rolle gpu-llamacpp setzt die UUID der P2000) EMBED_GPU=1 RERANK_GPU=1 CHAT_GPU=1