# spark-moe runtime env (GX10)
SPARK_MODEL_PATH=/path/to/glm-5.2-nvfp4
SPARK_MEM_AVAILABLE_GIB=auto
SPARK_OS_RESERVE_GIB=10
SPARK_RESIDENT_GIB=35
SPARK_SCRATCH_GIB=7
SPARK_KV_DTYPE=fp8
SPARK_MAX_CONTEXT=262144
SPARK_KV_BUDGET_GIB=auto
SPARK_MAX_NUM_SEQS=1          # KV slots + queue depth (sequential forward)
SPARK_QUALITY_TEMP=0          # parity CLI greedy decode
SPARK_BENCH_CONTEXT=0         # set 1 to run 8K-256K bench in run-tests
SPARK_SGLANG_PARITY=0         # set 1 + SGLANG_URL for live parity
SPARK_MIN_HOT_TIER_GIB=32
SPARK_HOT_TIER=auto
SPARK_HOT_TIER_GIB=auto
SPARK_EXPERT_BYTES_ESTIMATE=19922944
SPARK_KV_BLOCK_SIZE=256
SPARK_API_PORT=8000
SPARK_GEMM_CUDA=1
SPARK_GEMM_MARLIN_TC=1
SPARK_FORWARD_MOE_LAYERS=auto
SPARK_FORWARD_MAX_SEQ=auto
SPARK_MTP_FULL_LAYER=0
SPARK_PARITY_BASELINE=fixtures/spark-quality-golden.jsonl
SPARK_EXPERT_PREFETCH=1
SPARK_EXPERT_PREFETCH_ASYNC=0
SPARK_INDEX_SHARE=1
SPARK_ALLOW_STUB=0
SPARK_PREFILL_CHUNK=2048
SPARK_RUN_DOCKER_TESTS=0