MLEvolve/run_single_task.sh at main · InternScience/MLEvolve · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
#!/bin/bash
# Run MLEvolve on a single competition task.
# Usage: bash run_single_task.sh <EXP_ID> <DATASET_DIR> [SERVER_ID]
set -x

EXP_ID=${1:?Usage: bash run_single_task.sh <EXP_ID> <DATASET_DIR> [SERVER_ID]}
dataset_dir=${2:?Usage: bash run_single_task.sh <EXP_ID> <DATASET_DIR> [SERVER_ID]}
SERVER_ID=${3:-111}

# ── Proxy (uncomment & fill in if behind a corporate firewall) ──
# export http_proxy=http://YOUR_PROXY:PORT
# export https_proxy=http://YOUR_PROXY:PORT

ROOT="$(cd "$(dirname "$0")" && pwd)"
cd "$ROOT"

# ── Launch the local grading (format-validation) server ──
export DATASET_DIR="${dataset_dir}"
bash "$ROOT/launch_server.sh" "${SERVER_ID}"

BASE_PORT=5005
GRADING_SERVER_PORT=$((BASE_PORT + SERVER_ID))
export GRADING_SERVER_PORT

echo "Waiting for grading server on port ${GRADING_SERVER_PORT} ..."
MAX_WAIT=30
WAITED=0
while [ $WAITED -lt $MAX_WAIT ]; do
    if curl -s "http://127.0.0.1:${GRADING_SERVER_PORT}/health" > /dev/null 2>&1; then
        echo "Grading server ready (port ${GRADING_SERVER_PORT})."
        break
    fi
    sleep 1
    WAITED=$((WAITED + 1))
done
if [ $WAITED -ge $MAX_WAIT ]; then
    echo "Warning: grading server may not be ready yet, proceeding anyway ..."
fi

# ── Experiment settings ──
MEMORY_INDEX=0
start_cpu=0
CPUS_PER_TASK=21
TIME_LIMIT_SECS=43200           # 12 hours

export MEMORY_INDEX
format_time() {
  local t=$1
  echo "$((t/3600))hrs $(((t%3600)/60))mins $((t%60))secs"
}
export TIME_LIMIT=$(format_time $TIME_LIMIT_SECS)
export STEP_LIMIT=500

TIMESTAMP=$(date +"%Y%m%d_%H%M%S")
CLOSEST_EXP_NAME="${TIMESTAMP}_${EXP_ID}"

# ── HuggingFace cache (optional, point to a shared directory) ──
# export HF_ENDPOINT=https://huggingface.co
# export HF_DATASETS_CACHE=/path/to/hf_cache
# export HF_MODELS_CACHE=/path/to/hf_cache
# export HUGGINGFACE_HUB_CACHE=/path/to/hf_cache
# export TRANSFORMERS_CACHE=/path/to/hf_cache


# ── Run the main agent loop ──
CUDA_VISIBLE_DEVICES=$MEMORY_INDEX timeout --foreground --signal=TERM --kill-after=10s "${TIME_LIMIT_SECS}s" python run.py \
  exp_id="${EXP_ID}" \
  dataset_dir="${dataset_dir}" \
  data_dir="${dataset_dir}/${EXP_ID}/prepared/public" \
  desc_file="${dataset_dir}/${EXP_ID}/prepared/public/description.md" \
  exp_name="${EXP_ID}" \
  start_cpu_id="${start_cpu}" \
  cpu_number="${CPUS_PER_TASK}"

RUN_EXIT=$?

if [ $RUN_EXIT -eq 124 ]; then
  echo "Timed out after $TIME_LIMIT"
elif [ $RUN_EXIT -eq 130 ]; then
  echo "Interrupted."
  exit 130
elif [ $RUN_EXIT -ne 0 ]; then
  echo "Run failed with exit code: $RUN_EXIT"
fi

# ── Post-processing: ensemble top solutions ──
echo "Running submission fusion ..."
python utils/submission_fusion_utils.py \
  --task_id "${EXP_ID}" \
  --exp_name "${CLOSEST_EXP_NAME}"