config.conf.example

# sparkcli configuration
# Copy to ~/.sparkcli/config.conf and adjust for your setup.
# Run 'sparkcli doctor' to verify your configuration.

# Host-side HuggingFace cache directory.
# Files here are owned by root (written by Docker) — don't change this
# unless you know what you're doing; mismatched ownership breaks vLLM.
SPARKCLI_HF_CACHE=~/.cache/huggingface

# Docker image to use for vLLM (must have huggingface-cli + vLLM installed).
SPARKCLI_VLLM_IMAGE=sparky-vllm:26.02

# Path to the directory containing your vLLM Dockerfile.
# Required only if you use 'sparkcli update' to rebuild the image.
# SPARKCLI_VLLM_BUILD_DIR=~/homelab/docker/vllm

# Port vLLM listens on (host-side). OpenWebUI should point here.
SPARKCLI_PORT=8000

# GPU memory fraction vLLM may use. 0.80 leaves headroom for the OS.
# On a DGX Spark with 128GB unified memory you can push this to 0.90+
# if you're running large models.
SPARKCLI_GPU_UTIL=0.80

# Optional: HuggingFace token for gated models.
# Can also be set as HF_TOKEN or HUGGING_FACE_HUB_TOKEN in the environment.
# HF_TOKEN=hf_...