# Profile: a (weak) GPU box. Use with the GPU override:
#   cp profiles/gpu.env .env
#   docker compose -f docker-compose.yml -f docker-compose.gpu.yml up -d
#
# NGL=99 offloads all model layers; lower it (e.g. 20) if VRAM < model size.
# The BF16 mmproj stays on CPU (--no-mmproj-offload in the llama-swap macro).
HF_TOKEN=

NGL=99
THREADS=8
CTX=32768
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)

# CUDA images (override base):
LLAMASWAP_IMAGE=ghcr.io/mostlygeek/llama-swap:cuda
TEI_IMAGE=ghcr.io/huggingface/text-embeddings-inference: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 (gpu)