#!/usr/bin/env bash
# Score several profiles at once, one worker per card.
#
# The serial scorer manages roughly one video per minute, so the 21 profiles of
# a full batch cost most of a day on their own.  Sharding by profile is safe
# because each profile writes only into its own eval root, and `run_eval.sh`
# skips anything already carrying a completion marker, so workers never contend
# and a re-run is idempotent.
set -uo pipefail
cd /home/n2501108a/KV/experiments/rope_reindex_allmethods_20260824

GPUS=${LINGBOT_EVAL_GPUS:?usage: LINGBOT_EVAL_GPUS=1,3,4,5 run_eval_parallel.sh TAG...}
[[ $# -gt 0 ]] || { echo "pass at least one profile tag" >&2; exit 2; }

IFS=',' read -r -a cards <<<"$GPUS"
log() { echo "[$(date '+%m-%d %H:%M:%S')] $*"; }

# Round-robin so that a long profile does not leave one card idle at the end.
declare -A shard
index=0
for tag in "$@"; do
  card=${cards[$((index % ${#cards[@]}))]}
  shard[$card]="${shard[$card]:-} $tag"
  index=$((index + 1))
done

pids=()
for card in "${cards[@]}"; do
  tags=${shard[$card]:-}
  [[ -n "$tags" ]] || continue
  log "GPU $card <-$tags"
  LINGBOT_EVAL_GPU="$card" LINGBOT_RUNS_DIR="${LINGBOT_RUNS_DIR:-runs}" \
    ./run_eval.sh $tags >"logs/eval_gpu${card}.log" 2>&1 &
  pids+=($!)
done

status=0
for pid in "${pids[@]}"; do
  wait "$pid" || status=1
done
log "all shards finished (status $status)"
exit "$status"