# Copy to .env and edit.
HF_TOKEN=hf_your_token_here

# Server
HOST=0.0.0.0
PORT=8000
SERVED_MODEL_NAME=qwen36-35b-a3b-fp8-262k

# Model
MODEL_ID=Qwen/Qwen3.6-35B-A3B-FP8

# DGX Spark / GB10 image.
VLLM_IMAGE=vllm/vllm-openai:cu130-nightly-aarch64
VLLM_REQUIRE_MIN_VERSION=0.19.0
VLLM_FAIL_IF_VERSION_TOO_OLD=0

# 262K context settings tested on DGX Spark / GB10.
MAX_MODEL_LEN=262144
MAX_NUM_BATCHED_TOKENS=8192
MAX_NUM_SEQS=1
GPU_MEMORY_UTILIZATION=0.82

# Serving settings.
KV_CACHE_DTYPE=fp8
ATTENTION_BACKEND=flashinfer
LOAD_FORMAT=fastsafetensors

# Qwen3.6 behavior.
LANGUAGE_MODEL_ONLY=1
REASONING_PARSER=qwen3

# Tool calling.
ENABLE_AUTO_TOOL_CHOICE=1
TOOL_CALL_PARSER=qwen3_xml

# Thinking behavior.
PRESERVE_THINKING=1

# CPU / stability knobs for Grace CPU + Triton/PyTorch.
OMP_NUM_THREADS=16
VLLM_CPU_OMP_THREADS_BIND=1
TORCHINDUCTOR_MAX_AUTOTUNE=0
TRITON_MAX_AUTOTUNE=0

# MTP is OFF by default.
ENABLE_MTP=0
MTP_METHOD=qwen3_next_mtp
MTP_NUM_SPECULATIVE_TOKENS=2

# Optional extra vLLM flags, appended at the end.
VLLM_EXTRA_ARGS=