#!/usr/bin/env bash
# MTP (multi-token-prediction) speculative-decode sweep for the latency report.
# Serves the BF16 model three ways — no speculation (baseline), MTP with 1
# speculative token, MTP with 2 — and captures the same page subset through the
# OCR API for each. The model ships 1 nextn-predict layer, so MTP=1 is the
# expected-valid setting and MTP=2 is exploratory.
#
# Run INSIDE the bench container:  bash scripts/quant/tune_mtp.sh
# Requires: api.py running on :5002 pointed at :8000; compat libs on path.
set -uo pipefail

MODEL="${MODEL:-datalab-to/surya-ocr-2}"
OCR_URL="http://127.0.0.1:5002/v1/api/ai/suya_ocr_vllm/"
N_PAGES="${N_PAGES:-5}"
OUT_ROOT="results/quant/mtp"
LOG=/tmp/tune_mtp.log
export LD_LIBRARY_PATH="/usr/local/cuda/compat:${LD_LIBRARY_PATH:-}"

mapfile -t ALL < <(sed '/^#/d;/^$/d' eval_set/manifest.txt)
IMAGES=("${ALL[@]:0:$N_PAGES}")

# config_name | extra vLLM args. The speculative-config JSON is written compact
# (no spaces) so it survives word-splitting as a single argv token.
CONFIGS=(
  "baseline|--enable-prefix-caching"
  "mtp1|--enable-prefix-caching --speculative-config {\"method\":\"mtp\",\"num_speculative_tokens\":1}"
  "mtp2|--enable-prefix-caching --speculative-config {\"method\":\"mtp\",\"num_speculative_tokens\":2}"
)

base_args() {
  echo "--host 127.0.0.1 --port 8000 --model $MODEL --served-model-name $MODEL \
--max-model-len 18000 --max-num-seqs 16 --gpu-memory-utilization 0.85 --no-enforce-eager \
--mm-processor-kwargs {\"min_pixels\":3136,\"max_pixels\":6291456}"
}

stop_server() {
  local pid
  pid=$(ss -ltnp 2>/dev/null | grep ":8000 " | grep -oP 'pid=\K[0-9]+' | head -1)
  [ -n "$pid" ] && kill -9 "$pid" 2>/dev/null
  pkill -9 -f "vllm.entrypoints" 2>/dev/null
  # EngineCore outlives the API server and holds the GPU; kill every remaining
  # CUDA compute process, then poll until the memory is actually freed.
  for p in $(nvidia-smi --query-compute-apps=pid --format=csv,noheader 2>/dev/null); do
    kill -9 "$p" 2>/dev/null
  done
  for _ in $(seq 1 30); do
    used=$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | head -1)
    [ "${used:-99999}" -lt 2000 ] && break
    sleep 2
  done
}

for entry in "${CONFIGS[@]}"; do
  name="${entry%%|*}"
  extra="${entry#*|}"
  echo "=== mtp config: $name ($extra) ===" | tee -a "$LOG"
  stop_server
  # shellcheck disable=SC2046
  nohup python3 -m vllm.entrypoints.openai.api_server $(base_args) $extra \
    > "/tmp/mtp_${name}.log" 2>&1 &
  ok=0
  for _ in $(seq 1 96); do
    if curl -fs http://127.0.0.1:8000/health >/dev/null 2>&1; then ok=1; break; fi
    if grep -qiE "Engine core initialization failed|ValueError|RuntimeError|Traceback" "/tmp/mtp_${name}.log" 2>/dev/null; then break; fi
    sleep 5
  done
  if [ "$ok" -ne 1 ]; then
    echo "$name: FAILED to start (see /tmp/mtp_${name}.log)" | tee -a "$LOG"
    continue
  fi
  python3 -m scripts.quant.capture --url "$OCR_URL" --out-dir "$OUT_ROOT/$name" "${IMAGES[@]}" \
    >> "$LOG" 2>&1
  echo "$name: captured $(ls "$OUT_ROOT/$name"/*.json 2>/dev/null | wc -l) pages" | tee -a "$LOG"
done
stop_server
echo "MTP_DONE" | tee -a "$LOG"