# Engine command line as it ran (from /proc/<pid>/cmdline on lm-server during the measurement, API key not part of it),
# and the environment the launcher gave it. The server layer (serve/server.py) starts the engine; it adds --serve,
# --layer-split 23 and --batch 2 (from "layer_split": "23" and "parallel": 2 in strata-q4xl-prod.json). The key goes
# in through the STRATA_API_KEY environment variable (systemd credential) and is not shown here.
# --pcie-frac 0.1 is in the command line as it ran; whether it changes anything on this box was not tested for this report.

/opt/strata-bmtp/build/strata-w7 --serve \
  --pack /opt/strata-ssd/packs/ud-q4_k_xl \
  --native /opt/strata-ssd/models/UD-Q4_K_XL/Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf \
  --expert-profile /opt/strata-bmtp/data/expert-profile.bin \
  --expert-cache auto \
  --spec 4 \
  --spec-min-p 0.5 \
  --mtp /opt/strata-ssd/mtp/rt \
  --max-context 262144 \
  --kv int8 \
  --kv-resident 32768 \
  --kv-pool-tokens 524288 \
  --resident-experts \
  --ple-io direct \
  --trim-stage-weights \
  --vram-reserve-mib 1500 \
  --vision \
  --batch-mtp \
  --pcie-frac 0.1 \
  --expert-profile-save /opt/strata-bmtp/data/expert-profile.learned.bin \
  --adapt-every 2 \
  --adapt-async 1 \
  --conversation-cache-mib 4096 \
  --conversation-cache-slots 3 \
  --memory-limit-mib 101376 \
  --aux-cpus auto \
  --prefill auto:16384 \
  --conversation-cache-min-free-mib 2048 \
  --layer-split 23 --batch 2

# environment (from the "env" block of strata-q4xl-prod.json; CUDA_DEVICE_ORDER=PCI_BUS_ID, CUDA_VISIBLE_DEVICES=0,1,
# LD_LIBRARY_PATH=/usr/local/cuda-13.3/bin:/usr/local/cuda-13.3/lib64 were also set)
STRATA_LOOKAHEAD=0
STRATA_SPLIT_TIMING=1
STRATA_ENGINE_READY_S=0
STRATA_SPLIT_MTP_BATCH=1
STRATA_PREFILL_RING=384
STRATA_PREFILL_EQUAL=1
