#!/usr/bin/env bash # Validated on the fixed 128+128-token office regression groups. This selects # the optimized arithmetic profile; it does not change CPU/NPU governors. set -euo pipefail if (( $# < 2 )); then echo "Usage: bash tools/run_optimized.sh {benchmark|generate|serve} MODEL.l3r [ARGS...]" >&2 exit 2 fi engine_root=$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." && pwd) engine_binary=${LING3_BINARY:-$engine_root/build-rk3588/ling3-rknn} [[ -x $engine_binary ]] || { echo "RKNN binary not found: $engine_binary" >&2; exit 2; } export LING3_PREWARM_EXPERTS=1 LING3_EXPERT_BALANCED=1 LING3_EXPERT_ZERO_COPY=1 export LING3_GDN_CPU_PREFILL=1 LING3_GDN_CPU_FP32_STATE=1 export LING3_GDN_CPU_DECODE=1 LING3_GDN_FULL_FP32=1 export LING3_MLA_SIMD=1 LING3_VECTOR_MATH=1 export LING3_GDN_PREFILL_DIR=${LING3_GDN_PREFILL_DIR:-$engine_root/gdn-prefill-t16} unset LING3_PREFILL_W4A4 LING3_EXPERT_ALL_CORES unset LING3_DISABLE_BATCH_INPUT_REUSE LING3_DISABLE_PARALLEL_GATHER exec "$engine_binary" "$@"