#!/bin/bash
# Score the four PostTrainBench base models on the offline capability probes.
#
#   sbatch --array=0-3 scripts/ptb_probes.sbatch
#
# One array task per model, one GPU each: the probes are a few thousand short
# log-likelihood passes, so this is minutes of compute once the weights load.
#
# Scored under both pmi and letter. Plain mean-token-logprob ("norm") is
# deliberately not the default: a positive control showed it collapses onto
# whichever option has the highest unconditional likelihood, putting even a
# 27B instruct model below chance.
# Results land in analysis/out/probe_results/<model>.json; summarise them with
#   python analysis/summarize_probes.py
# OUT_TAG suffixes the filename. Set it for any re-run that scores fewer
# probes or fewer schemes than the file already there, or the narrower run
# silently replaces a broader one:
#   OUT_TAG=_variants SCORING=pmi VARIANTS=base,swap sbatch --array=0-3 ...
#SBATCH --job-name=ptb-probes
#SBATCH --partition=h100
#SBATCH --account=midpri
#SBATCH --qos=midpri
#SBATCH --gres=gpu:1
#SBATCH --cpus-per-task=12
#SBATCH --mem=128G
#SBATCH --time=01:00:00
#SBATCH --output=logs/ptb-probes-%A_%a.out

set -euo pipefail
# The login pod's proxy is not routable from a compute node, and an inherited
# http_proxy makes every HF call hang before falling back to the local cache.
unset http_proxy https_proxy HTTP_PROXY HTTPS_PROXY no_proxy NO_PROXY X2P_PROXY_URL

REPO=/home/jiaruiliu/local/git-repos/auto-impl
cd "$REPO"

export HF_HOME=/lustre-storage/fsx_2/user/jiaruiliu/ptb-data/hf_cache
export HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 HF_DATASETS_OFFLINE=1
export TOKENIZERS_PARALLELISM=false
# A compile cache shared across concurrent engines on one node triggers an
# Errno 116 storm on Lustre; the judge pool learned this the hard way.
export VLLM_CACHE_ROOT="/tmp/vllm-cache-${SLURM_JOB_ID}-${SLURM_ARRAY_TASK_ID:-0}"
mkdir -p "$VLLM_CACHE_ROOT"

MODELS=(
  "Qwen/Qwen3-4B-Base"
  "Qwen/Qwen3-1.7B-Base"
  "HuggingFaceTB/SmolLM3-3B-Base"
  "google/gemma-3-4b-pt"
)
REPO_ID="${MODELS[${SLURM_ARRAY_TASK_ID:-0}]}"
NAME="${REPO_ID//\//__}"

# Resolve the local snapshot rather than passing the repo id: offline mode
# still consults the hub index for a bare id and fails on some layouts.
SNAP=$(ls -d "$HF_HOME/hub/models--${REPO_ID//\//--}/snapshots/"*/ 2>/dev/null | head -1)
if [[ -z "$SNAP" ]]; then
    echo "No local snapshot for $REPO_ID under $HF_HOME/hub" >&2
    exit 1
fi
echo "model=$REPO_ID snapshot=$SNAP"
nvidia-smi -L

"$HOME/.venv-vllm-serve/bin/python" analysis/run_probes.py \
    --model "$SNAP" \
    --name "$REPO_ID" \
    --out "analysis/out/probe_results/${NAME}${OUT_TAG:-}.json" \
    --max-model-len "${MAX_MODEL_LEN:-4096}" \
    --gpu-util "${GPU_UTIL:-0.85}" \
    --scoring "${SCORING:-pmi,letter}" \
    --variants "${VARIANTS:-base}"