.env.example

# Tailnet exposure
TAILSCALE_IP=100.xxx.xxx.xxx
VLLM_PORT=8000
VLLM_LOCAL_PORT=8001
OPENAI_API_KEY=replace-me-with-a-long-random-token

# Model selection
MODEL_ID=google/gemma-4-31B-it
SERVED_MODEL_NAME=gemma-4-31b-it

# Runtime sizing
TENSOR_PARALLEL_SIZE=1
MAX_MODEL_LEN=32768
GPU_MEMORY_UTILIZATION=0.93
SHM_SIZE=32g

# Cache / offline behavior
HF_CACHE_DIR=/home/anthonywu/.cache/huggingface
HF_HUB_OFFLINE=1
TRANSFORMERS_OFFLINE=1
VLLM_NO_USAGE_STATS=1
DO_NOT_TRACK=1

# Gemma 4 defaults
# Leave empty on DGX Spark to keep multimodal image and audio support enabled.
VLLM_LANGUAGE_ONLY_FLAG=
VLLM_EXTRA_ARGS=--async-scheduling --kv-cache-dtype fp8 --limit-mm-per-prompt '{"image":4,"audio":1}' --max-num-seqs 64

# Container image
VLLM_IMAGE=vllm/vllm-openai:gemma4-cu130