# Profile: a bigger CPU server (>= 16 GiB RAM, many cores).
# Default model 12B (audio+image+text). Set THREADS to the server's PHYSICAL cores.
HF_TOKEN=

NGL=0
THREADS=16            # <-- set to physical core count of the server
CTX=16384
DEFAULT_MODEL=gemma-12b

TEI_EMBED_MODEL=BAAI/bge-m3
TEI_EMBED_POOLING=cls
TEI_RERANK_MODEL=BAAI/bge-reranker-v2-m3
TEI_DTYPE=float16            # float16 -> lighter Candle backend on CPU (saves RAM)

LLAMASWAP_IMAGE=ghcr.io/mostlygeek/llama-swap:cpu
TEI_IMAGE=ghcr.io/huggingface/text-embeddings-inference:cpu-1.9

WEBUI_PORT=3000
LLAMASWAP_PORT=8080
TEI_EMBED_PORT=8081
TEI_RERANK_PORT=8082
OPENAI_API_KEY=sk-local
TEI_API_KEY=x
WEBUI_NAME=Gemma 4 Local (server)