#!/bin/bash
# Chain I (GPU4): Instruct 8B + Instruct int4 twin on MM-Vet (n=30, seed42) + MathVista bridge
S=/testessfs10/users/zeyu.zhang/wangyu_ssd
ST=$S/stage0
PY=$S/vllm_env/bin/python
cd $ST
export PYTHONPATH=$ST PYTHONDONTWRITEBYTECODE=1
export PATH=$S/vllm_env/bin:$PATH
export HF_HOME=$S/hf_cache HF_ENDPOINT=https://hf-mirror.com
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
MI=Qwen/Qwen3-VL-8B-Instruct
DI=cyankiwi/Qwen3-VL-8B-Instruct-AWQ-4bit
CMV="--parquet mmvet.parquet --pids-from mmvet_pids.jsonl --n 30 --max-tokens 512 --min-tokens 64 --gpu-mem-util 0.60 --max-model-len 3584"
CBR="--parquet testmini.parquet --pids-from res05_qwen_thinking.jsonl --n 20 --max-tokens 1024 --min-tokens 64 --gpu-mem-util 0.60 --max-model-len 3584"
SMK="--parquet mmvet.parquet --pids-from mmvet_pids.jsonl --n 4 --max-tokens 128 --min-tokens 64 --gpu-mem-util 0.60 --max-model-len 3584"
B=$S/stage1
rm -f $B/ogi_*.jsonl $B/bri_*.jsonl $B/DONE_ogi $B/FAIL_ogi

gate() {
  A=$(grep -o "Draft acceptance rate: [0-9.]*" "$1" | tail -1 | sed "s/.*: //")
  awk -v a="${A:-0}" -v t="$2" "BEGIN{exit (a>=t)?0:1}"
}

nohup bash -c "
echo \"from huggingface_hub import snapshot_download; snapshot_download(\\\"$DI\\\")\" > /tmp/ensure_iawq.py
$PY /tmp/ensure_iawq.py
$(declare -f gate)
CUDA_VISIBLE_DEVICES=4 $PY stage1_spec_smoke.py --mode spec --model $MI --draft $DI --gamma 6 $SMK --out $B/ogi_smoke.jsonl > $B/logs/ogi_smoke.log 2>&1
if gate $B/logs/ogi_smoke.log 40; then EAG=; else EAG=--enforce-eager; touch $B/FAIL_ogi; fi
CUDA_VISIBLE_DEVICES=4 $PY stage1_spec_smoke.py --mode vanilla --model $MI $CMV \$EAG --out $B/ogi_van1.jsonl > $B/logs/ogi_van1.log 2>&1
CUDA_VISIBLE_DEVICES=4 $PY stage1_spec_smoke.py --mode spec --model $MI --draft $DI --gamma 6 $CMV \$EAG --out $B/ogi_strict_g6.jsonl > $B/logs/ogi_strict_g6.log 2>&1
VLLM_SPEC_RELAX_LOGTAU=0.693 CUDA_VISIBLE_DEVICES=4 $PY stage1_spec_smoke.py --mode spec --model $MI --draft $DI --gamma 6 $CMV \$EAG --out $B/ogi_relax_g6.jsonl > $B/logs/ogi_relax_g6.log 2>&1
CUDA_VISIBLE_DEVICES=4 $PY stage1_spec_smoke.py --mode vanilla --model $MI $CMV \$EAG --out $B/ogi_vanm.jsonl > $B/logs/ogi_vanm.log 2>&1
CUDA_VISIBLE_DEVICES=4 $PY stage1_spec_smoke.py --mode spec --model $MI --draft $DI --gamma 4 $CMV \$EAG --out $B/ogi_strict_g4.jsonl > $B/logs/ogi_strict_g4.log 2>&1
VLLM_SPEC_RELAX_LOGTAU=0.693 CUDA_VISIBLE_DEVICES=4 $PY stage1_spec_smoke.py --mode spec --model $MI --draft $DI --gamma 4 $CMV \$EAG --out $B/ogi_relax_g4.jsonl > $B/logs/ogi_relax_g4.log 2>&1
CUDA_VISIBLE_DEVICES=4 $PY stage1_spec_smoke.py --mode vanilla --model $MI $CMV \$EAG --out $B/ogi_van2.jsonl > $B/logs/ogi_van2.log 2>&1
CUDA_VISIBLE_DEVICES=4 $PY stage1_spec_smoke.py --mode vanilla --model $MI $CBR \$EAG --out $B/bri_van1.jsonl > $B/logs/bri_van1.log 2>&1
CUDA_VISIBLE_DEVICES=4 $PY stage1_spec_smoke.py --mode spec --model $MI --draft $DI --gamma 6 $CBR \$EAG --out $B/bri_strict_g6.jsonl > $B/logs/bri_strict_g6.log 2>&1
CUDA_VISIBLE_DEVICES=4 $PY stage1_spec_smoke.py --mode vanilla --model $MI $CBR \$EAG --out $B/bri_van2.jsonl > $B/logs/bri_van2.log 2>&1
touch $B/DONE_ogi
" > /dev/null 2>&1 &
disown
echo LAUNCHED_INSTRUCT