HF_TOKEN=

# Container image settings.
VLLM_IMAGE=vllm/vllm-openai:gemma
IMAGE_NAME=diffusiongemma-26b-a4b-nvfp4-vllm:latest
CONTAINER_NAME=diffusiongemma-26b-a4b-nvfp4

# Model and API settings.
MODEL_ID=nvidia/diffusiongemma-26B-A4B-it-NVFP4
SERVED_MODEL_NAME=diffusiongemma-26b-a4b-nvfp4

HOST=0.0.0.0
PORT=8000

# vLLM capacity settings. Tune these down if the model does not fit.
MAX_MODEL_LEN=262144
MAX_NUM_SEQS=4
MAX_NUM_BATCHED_TOKENS=32768
GPU_MEMORY_UTILIZATION=0.85

# DiffusionGemma/vLLM runtime settings.
VLLM_USE_V2_MODEL_RUNNER=1
ATTENTION_BACKEND=TRITON_ATTN
ENABLE_AUTO_TOOL_CHOICE=1
TOOL_CALL_PARSER=gemma4
REASONING_PARSER=gemma4
OVERRIDE_GENERATION_CONFIG='{"max_new_tokens": null}'
CHAT_TEMPLATE_KWARGS='{"enable_thinking":true}'

# Optional compatibility switches.
DTYPE=auto
KV_CACHE_DTYPE=auto
QUANTIZATION=
VLLM_EXTRA_ARGS=

# Docker Compose runtime settings.
RESTART_POLICY=unless-stopped
MEM_LIMIT=112g
MEMSWAP_LIMIT=112g