code/slurm_logs/mega_pipeline_v1.sbatch

#!/bin/bash
#SBATCH --job-name=vl
#SBATCH --partition=gpu-large
#SBATCH --qos=batch-long
#SBATCH --gres=gpu:1
#SBATCH --cpus-per-task=4
#SBATCH --mem=80G
#SBATCH --time=12:00:00
#SBATCH --output=/weka/s225250685/mats-tist/slurm_logs/pipeline_v1_%j.out

# ============================================================
# Pipeline v1 — full pipeline with thanhdath Llama-3B planner
#
# Stage A: 1-stage K=8 oracle rollout (thanhdath Llama-3B planner)
# Stage B: Build ORPO data (val-sel, val-cond, exec-error fixer)
# Stage C: Train validators (Qwen 0.5B ORPO) + exec-error fixer (Qwen 1.5B ORPO)
# Stage D: 2-stage rollout (planner + gated exec-error fixer)
# Stage E: Compute oracle + metrics
# Stage F: Build selector training data + train selector (Qwen 3B SFT)
# Stage G: Final EX evaluation with selector
# ============================================================

set -u
cd /weka/s225250685/mats-tist

export HF_HOME=/weka/s225250685/Huggingface
export HF_HUB_CACHE=/weka/s225250685/Huggingface/hub
export DB_EXEC_API_DISABLE=1
export PYTHONNOUSERSITE=1
export NO_PROXY=localhost,127.0.0.1
export PYTHONPATH=/weka/s225250685/mats-tist
export TOKENIZERS_PARALLELISM=false

PY=/weka/s225250685/conda-envs/handbook/bin/python
VLLM=/weka/s225250685/conda-envs/handbook/bin/vllm
ACCEL=/weka/s225250685/conda-envs/handbook/bin/accelerate
AH=/weka/s225250685/mats-tist/alignment-handbook

THANHDATH_PLANNER=/weka/s225250685/Huggingface/hub/models--thanhdath--orpo-llama-3b-iter-2-bird-planner/snapshots/8171b8585a306709996796b86de19c3dd39a910c

VAL_SEL=$AH/output/validator-sel-v1-qwen-orpo
VAL_COND=$AH/output/validator-cond-v1-qwen-orpo
FIXER=$AH/output/fixer-v1-qwen-orpo
SELECTOR=$AH/output/selector-v1-qwen-sft

LOG=/weka/s225250685/mats-tist/slurm_logs/pipeline_v1_${SLURM_JOB_ID}.log
: > "$LOG"

nvidia-smi --query-gpu=name,memory.total --format=csv,noheader | tee -a "$LOG"

kill_vllm() {
    pkill -9 -f "vllm serve" 2>/dev/null || true
    pkill -9 -f "VLLM::EngineCore" 2>/dev/null || true
    sleep 5
}
trap kill_vllm EXIT

wait_url() {
    for i in {1..180}; do
        curl --noproxy '*' -fs "$1" >/dev/null 2>&1 && return 0
        sleep 5
    done
    echo "TIMEOUT waiting for $1" | tee -a "$LOG"
    return 1
}

##############################################
# STAGE A: 1-stage oracle rollout with thanhdath planner
##############################################
echo "==== [A] launching thanhdath Llama-3B planner ====" | tee -a "$LOG"
kill_vllm

$VLLM serve "$THANHDATH_PLANNER" --served-model-name planner --port 8100 \
    --dtype bfloat16 --gpu-memory-utilization 0.85 \
    --enforce-eager --max-model-len 8192 > "${LOG}.planner" 2>&1 &
wait_url http://localhost:8100/v1/models && echo "  planner READY" | tee -a "$LOG"

# Do NOT rm -f — script auto-resumes from existing output (skips already-processed questions)
ORACLE_OUT=eval_results/pipeline_v1_K8_1stage_thanhdath_bird_dev.jsonl

echo "==== [A] K=8 oracle rollout (Llama-3B, no V+F) ====" | tee -a "$LOG"
$PY scripts/run_pipeline_rollouts.py \
    --input_file data/sft_bird_with_evidence_dev_text2sql.json \
    --output_file "$ORACLE_OUT" \
    --planner_host http://localhost:8100 \
    --planner_format llama3 \
    --validator_host none \
    --fixer_host none \
    --K 8 --K_val 1 --K_fix 1 \
    --temperature 1.0 --top_p 0.9 \
    --max_planner_tokens 1024 \
    --max_questions -1 --n_threads 4 2>&1 | tee -a "$LOG"

echo "==== [A] oracle metrics ====" | tee -a "$LOG"
$PY scripts/compute_bestofn_metrics.py "$ORACLE_OUT" pipeline_v1_1stage 2>&1 | tee -a "$LOG"

kill_vllm

# STAGES B-F REMOVED: validators/fixers need SFT first with GRIFFITH schema
# (built from BIRD-TRAIN rollouts, not BIRD-DEV), then optionally ORPO on top.
# These are handled by separate jobs (mega_valfix_sft_griffith.sbatch).
echo "==== ALL_DONE (oracle only — val/fix training in separate job) ====" | tee -a "$LOG"
exit 0

##############################################
# STAGE B (DISABLED): Build ORPO data for validators and exec-error fixer
##############################################
echo "==== [B] building ORPO training data from rollout ====" | tee -a "$LOG"

# Build validator ORPO data from the oracle rollout
$PY - << 'PYEOF' 2>&1 | tee -a "$LOG"
import json, os, random, re
from datasets import Dataset, DatasetDict

ROOT = "/weka/s225250685/mats-tist"
os.chdir(ROOT)

ROLLOUT = "eval_results/pipeline_v1_K8_1stage_thanhdath_bird_dev.jsonl"
VAL_SEL_INSTR = ("You are a SQL SELECT-clause critique agent. Output ONE critique section "
                 "<select>...</select> analysing the SELECT clause of the SQL query below; "
                 "do NOT output any SQL. Use 'None' if the SELECT clause looks correct.")
VAL_COND_INSTR = ("You are a SQL CONDITION critique agent. Output ONE critique section "
                  "<condition>...</condition> analysing the WHERE/HAVING/CASE-WHEN conditions "
                  "of the SQL query below; do NOT output any SQL. Use 'None' if the conditions look correct.")

def schema_str(sample):
    return str(sample.get("schema_sequence") or sample.get("schema") or "")

def build_val_prompt(instr, schema, question, evidence, sql, exec_result):
    return (instr + "\n\ndatabase schema:\n" + schema +
            "\n\nQuestion: " + question +
            "\nExternal knowledge: " + (evidence or "None") +
            "\n\nGenerated SQL query: " + sql +
            "\n\nExecution response:\n" + exec_result + "\n\n")

def qwen_chat(prompt):
    return f"<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n"

def safe_exec(db_path, sql, timeout=5):
    import sqlite3, threading
    result=[None]; err=[None]
    def _run():
        try:
            conn = sqlite3.connect(db_path)
            conn.text_factory = lambda b: b.decode(errors="ignore")
            result[0] = conn.execute(sql).fetchmany(10)
            conn.close()
        except Exception as e:
            err[0] = str(e)
    t = threading.Thread(target=_run, daemon=True)
    t.start(); t.join(timeout)
    if t.is_alive(): return None, "TIMEOUT"
    return result[0], err[0]

random.seed(42)
with open(ROLLOUT) as f:
    lines = [json.loads(l) for l in f]

sel_pairs = []; cond_pairs = []

for ex in lines:
    db_path = ex["db_path"]
    schema = schema_str(ex)
    q = ex["question"]
    ev = ex.get("evidence", "")

    correct_trajs = [t for t in ex["trajectories"] if t.get("is_planner_correct")]
    wrong_trajs   = [t for t in ex["trajectories"] if not t.get("is_planner_correct")
                     and t.get("planner_exec_ok")]

    if not correct_trajs or not wrong_trajs:
        continue

    for ct in correct_trajs[:2]:
        sql_c = ct["planner_sql"]
        rows_c, err_c = safe_exec(db_path, sql_c)
        exec_r_c = f"OK. Result rows: {str(rows_c)[:300]}" if not err_c else f"Error: {err_c[:200]}"

        wt = random.choice(wrong_trajs)
        sql_w = wt["planner_sql"]
        rows_w, err_w = safe_exec(db_path, sql_w)
        exec_r_w = f"OK. Result rows: {str(rows_w)[:300]}" if not err_w else f"Error: {err_w[:200]}"

        # SELECT critique pairs
        prompt_c = qwen_chat(build_val_prompt(VAL_SEL_INSTR, schema, q, ev, sql_c, exec_r_c))
        prompt_w = qwen_chat(build_val_prompt(VAL_SEL_INSTR, schema, q, ev, sql_w, exec_r_w))
        # chosen = "None" (correct SQL), rejected = "INCORRECT: ..." (wrong SQL)
        sel_pairs.append({"prompt": prompt_c, "chosen": "<select>\nSELECT.\nNone\n</select>",
                          "rejected": "<select>\nSELECT.\nINCORRECT: SELECT clause is incorrect.\n</select>"})
        sel_pairs.append({"prompt": prompt_w, "chosen": "<select>\nSELECT.\nINCORRECT: SELECT clause produces wrong results.\n</select>",
                          "rejected": "<select>\nSELECT.\nNone\n</select>"})

        # CONDITION critique pairs
        prompt_c2 = qwen_chat(build_val_prompt(VAL_COND_INSTR, schema, q, ev, sql_c, exec_r_c))
        prompt_w2 = qwen_chat(build_val_prompt(VAL_COND_INSTR, schema, q, ev, sql_w, exec_r_w))
        cond_pairs.append({"prompt": prompt_c2, "chosen": "<condition>\nCONDITION.\nNone\n</condition>",
                           "rejected": "<condition>\nCONDITION.\nINCORRECT: WHERE/HAVING conditions are wrong.\n</condition>"})
        cond_pairs.append({"prompt": prompt_w2, "chosen": "<condition>\nCONDITION.\nINCORRECT: WHERE/HAVING conditions produce wrong results.\n</condition>",
                           "rejected": "<condition>\nCONDITION.\nNone\n</condition>"})

random.shuffle(sel_pairs); random.shuffle(cond_pairs)
n_sel = int(0.9 * len(sel_pairs))
n_cond = int(0.9 * len(cond_pairs))

DatasetDict({"train_dpo": Dataset.from_list(sel_pairs[:n_sel]),
             "test_dpo":  Dataset.from_list(sel_pairs[n_sel:])}).save_to_disk("data/hf_val_sel_v1_orpo")
print(f"val-sel ORPO: {n_sel} train + {len(sel_pairs)-n_sel} test")

DatasetDict({"train_dpo": Dataset.from_list(cond_pairs[:n_cond]),
             "test_dpo":  Dataset.from_list(cond_pairs[n_cond:])}).save_to_disk("data/hf_val_cond_v1_orpo")
print(f"val-cond ORPO: {n_cond} train + {len(cond_pairs)-n_cond} test")
PYEOF

# Build exec-error fixer ORPO data (reuse existing script with new rollout)
echo "  building exec-error fixer ORPO data..." | tee -a "$LOG"
$PY scripts/build_fixer_v2_execerr.py 2>&1 | tee -a "$LOG"
[ -L data/hf_fixer_v2_execerr_expanded ] || ln -sf hf_fixer_v2_execerr data/hf_fixer_v2_execerr_expanded

echo "==== [B] data build done ====" | tee -a "$LOG"

##############################################
# STAGE C: Train validators + exec-error fixer (ORPO)
##############################################
echo "==== [C] ORPO validator-sel (Qwen 0.5B) ====" | tee -a "$LOG"
cd $AH
PYTHONPATH=src/ ACCELERATE_LOG_LEVEL=info $ACCEL launch \
    --main_process_port 29700 \
    --config_file recipes/accelerate_configs/single_gpu0_local.yaml \
    scripts/run_orpo.py \
    recipes/scaleup-3stage/orpo-validator-sel-v1.yaml 2>&1 | tee -a "$LOG"
cd /weka/s225250685/mats-tist
[ -f "$VAL_SEL/config.json" ] || { echo "VAL_SEL ORPO FAILED"; exit 1; }

echo "==== [C] ORPO validator-cond (Qwen 0.5B) ====" | tee -a "$LOG"
cd $AH
PYTHONPATH=src/ ACCELERATE_LOG_LEVEL=info $ACCEL launch \
    --main_process_port 29701 \
    --config_file recipes/accelerate_configs/single_gpu0_local.yaml \
    scripts/run_orpo.py \
    recipes/scaleup-3stage/orpo-validator-cond-v1.yaml 2>&1 | tee -a "$LOG"
cd /weka/s225250685/mats-tist
[ -f "$VAL_COND/config.json" ] || { echo "VAL_COND ORPO FAILED"; exit 1; }

echo "==== [C] ORPO exec-error fixer (Qwen 1.5B) ====" | tee -a "$LOG"
cd $AH
PYTHONPATH=src/ ACCELERATE_LOG_LEVEL=info $ACCEL launch \
    --main_process_port 29702 \
    --config_file recipes/accelerate_configs/single_gpu0_local.yaml \
    scripts/run_orpo.py \
    recipes/scaleup-3stage/orpo-fixer-v1.yaml 2>&1 | tee -a "$LOG"
cd /weka/s225250685/mats-tist
[ -f "$FIXER/config.json" ] || { echo "FIXER ORPO FAILED"; exit 1; }

echo "==== [C] all models trained ====" | tee -a "$LOG"

##############################################
# STAGE D: 2-stage rollout (planner + gated exec-error fixer)
#   H200=141GB: planner(0.30)+val_sel(0.08)+val_cond(0.08)+fixer(0.15) = 0.61
##############################################
kill_vllm
echo "==== [D] launching 4 endpoints ====" | tee -a "$LOG"

$VLLM serve "$THANHDATH_PLANNER" --served-model-name planner --port 8100 \
    --dtype bfloat16 --gpu-memory-utilization 0.30 \
    --enforce-eager --max-model-len 8192 > "${LOG}.p" 2>&1 &
wait_url http://localhost:8100/v1/models && echo "  planner READY" | tee -a "$LOG"

$VLLM serve "$VAL_SEL" --served-model-name validator_sel --port 8101 \
    --dtype bfloat16 --gpu-memory-utilization 0.08 \
    --enforce-eager --max-model-len 6144 > "${LOG}.vs" 2>&1 &
wait_url http://localhost:8101/v1/models && echo "  validator-sel READY" | tee -a "$LOG"

$VLLM serve "$VAL_COND" --served-model-name validator_cond --port 8104 \
    --dtype bfloat16 --gpu-memory-utilization 0.08 \
    --enforce-eager --max-model-len 6144 > "${LOG}.vc" 2>&1 &
wait_url http://localhost:8104/v1/models && echo "  validator-cond READY" | tee -a "$LOG"

$VLLM serve "$FIXER" --served-model-name fixer --port 8102 \
    --dtype bfloat16 --gpu-memory-utilization 0.15 \
    --enforce-eager --max-model-len 4096 > "${LOG}.f" 2>&1 &
wait_url http://localhost:8102/v1/models && echo "  fixer READY" | tee -a "$LOG"

ROLLOUT2=eval_results/pipeline_v1_K8_2stage_thanhdath_val_fixer_bird_dev.jsonl
rm -f "$ROLLOUT2"

echo "==== [D] 2-stage K=8 rollout ====" | tee -a "$LOG"
$PY scripts/run_pipeline_rollouts.py \
    --input_file data/sft_bird_with_evidence_dev_text2sql.json \
    --output_file "$ROLLOUT2" \
    --planner_host http://localhost:8100 \
    --planner_format llama3 \
    --validator_host none \
    --validator_sel_host http://localhost:8101 \
    --validator_cond_host http://localhost:8104 \
    --fixer_host http://localhost:8102 \
    --fixer_gate_exec_ok \
    --K 8 --K_val 1 --K_fix 1 \
    --temperature 1.0 --top_p 0.9 \
    --max_planner_tokens 1024 --max_validator_tokens 384 --max_fixer_tokens 512 \
    --max_questions -1 --n_threads 4 2>&1 | tee -a "$LOG"

echo "==== [D] metrics ====" | tee -a "$LOG"
$PY scripts/compute_bestofn_metrics.py "$ROLLOUT2" pipeline_v1_2stage 2>&1 | tee -a "$LOG"

##############################################
# STAGE E: Build selector training data + train selector
##############################################
kill_vllm
echo "==== [E] building selector training data ====" | tee -a "$LOG"

$PY - << 'PYEOF' 2>&1 | tee -a "$LOG"
import json, os, random, sqlite3, threading
from datasets import Dataset, DatasetDict
from data_processing.planner import is_execution_correct
ROOT = "/weka/s225250685/mats-tist"
os.chdir(ROOT)

PROMPT_TMPL = (
    "You are a SQL correctness judge.\n"
    "Schema:\n{schema}\n\n"
    "Question: {question}\n"
    "External knowledge: {evidence}\n\n"
    "Candidate SQL:\n{sql}\n\n"
    "Execution result:\n{exec_result}\n\n"
    "Is this SQL correct for the question? Answer YES or NO."
)

def safe_exec(db_path, sql, timeout=5):
    result=[None]; err=[None]
    def _run():
        try:
            conn = sqlite3.connect(db_path)
            conn.text_factory = lambda b: b.decode(errors="ignore")
            result[0] = conn.execute(sql).fetchmany(5)
            conn.close()
        except Exception as e:
            err[0] = str(e)
    t = threading.Thread(target=_run, daemon=True)
    t.start(); t.join(timeout)
    if t.is_alive(): return None, "TIMEOUT"
    return result[0], err[0]

def qwen_chat(prompt):
    return f"<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n"

rows = []
random.seed(42)

for rollout_file in ["eval_results/pipeline_v1_K8_1stage_thanhdath_bird_dev.jsonl",
                     "eval_results/pipeline_v1_K8_2stage_thanhdath_val_fixer_bird_dev.jsonl"]:
    if not os.path.exists(rollout_file):
        continue
    with open(rollout_file) as f:
        for line in f:
            ex = json.loads(line)
            db_path = ex["db_path"]
            schema = str(ex.get("schema_sequence") or ex.get("schema") or "")
            q = ex["question"]
            ev = ex.get("evidence", "") or "None"
            gold_sql = ex["sql"]
            gold_res, gold_err = safe_exec(db_path, gold_sql)
            if gold_err: continue

            for t in ex["trajectories"]:
                sql = t.get("fixed_sql") or t.get("planner_sql") or ""
                if not sql.strip(): continue
                res, err = safe_exec(db_path, sql)
                if err:
                    exec_str = f"Error: {err[:200]}"
                else:
                    exec_str = f"OK. Rows preview: {str(res)[:300]}"
                label = "YES" if (not err and is_execution_correct(gold_res, res)) else "NO"
                prompt = PROMPT_TMPL.format(schema=schema[:3000], question=q,
                                            evidence=ev, sql=sql[:800], exec_result=exec_str[:300])
                rows.append({"prompt": qwen_chat(prompt), "completion": label,
                             "label": label, "db_id": ex.get("db_id","")})

random.shuffle(rows)
n = int(0.9 * len(rows))
DatasetDict({"train": Dataset.from_list(rows[:n]),
             "test":  Dataset.from_list(rows[n:])}).save_to_disk("data/hf_selector_v1")
print(f"selector data: {n} train + {len(rows)-n} test  (YES={sum(1 for r in rows if r['label']=='YES')}, NO={sum(1 for r in rows if r['label']=='NO')})")
PYEOF

echo "==== [E] training selector (Qwen2.5-Coder-3B) ====" | tee -a "$LOG"
$PY scripts/train_selector_v1.py \
    --base Qwen/Qwen2.5-Coder-3B-Instruct \
    --data data/hf_selector_v1 \
    --out "$SELECTOR" \
    --epochs 2 --lr 1e-5 --bs 1 --grad_accum 64 --max_len 4096 2>&1 | tee -a "$LOG"

[ -f "$SELECTOR/config.json" ] || { echo "SELECTOR TRAIN FAILED"; exit 1; }

##############################################
# STAGE F: Final evaluation with selector
##############################################
kill_vllm
echo "==== [F] launching selector ====" | tee -a "$LOG"
$VLLM serve "$SELECTOR" --served-model-name selector --port 8103 \
    --dtype bfloat16 --gpu-memory-utilization 0.85 \
    --enforce-eager --max-model-len 4096 > "${LOG}.sel" 2>&1 &
wait_url http://localhost:8103/v1/models && echo "  selector READY" | tee -a "$LOG"

for OUT in "$ORACLE_OUT" "$ROLLOUT2"; do
    label="$(basename $OUT .jsonl)_selV1"
    echo "  scoring $label" | tee -a "$LOG"
    $PY scripts/compute_bestofn_with_selector.py \
        "$OUT" "$label" --selector_host http://localhost:8103 --row_preview 2>&1 | tee -a "$LOG"
done

echo "==== ALL_DONE ====" | tee -a "$LOG"