#!/usr/bin/env bash
# Start llama-server for GLM-5.3 E192 UD-IQ2_M on one NVIDIA B200 (183 GB).
# Every setting can be overridden with an environment variable; extra arguments are passed to llama-server.
#   LLAMA_SERVER=./build/bin/llama-server MODEL_DIR=/path/to/model ./serve_b200.sh
#   DRY_RUN=1 ./serve_b200.sh        # print the command only
set -euo pipefail

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"

LLAMA_SERVER="${LLAMA_SERVER:-llama-server}"
MODEL_DIR="${MODEL_DIR:-$SCRIPT_DIR/..}"
MODEL="${MODEL:-$MODEL_DIR/GLM-5.3-UD-IQ2_M-E192-00001-of-00006.gguf}"
HOST="${HOST:-127.0.0.1}"
PORT="${PORT:-8080}"
CTX="${CTX:-102400}"            # shared KV pool in tokens; 102400 leaves ~1.7 GB free on a B200
NP="${NP:-10}"                  # parallel slots
SLOT_CTX="${SLOT_CTX:-68608}"   # per-slot limit: longest GPQA prompt (2821 tokens) + 65536 generated tokens

CMD=("$LLAMA_SERVER" -m "$MODEL" -ngl 99 -kvu -c "$CTX" --kv-unified-per-slot "$SLOT_CTX" -np "$NP" -bs
     --host "$HOST" --port "$PORT" --metrics "$@")

if [[ "${DRY_RUN:-0}" == "1" ]]; then
    printf '%q ' "${CMD[@]}"; echo
    exit 0
fi
exec "${CMD[@]}"