LLAMACPP_REF=b6000
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