#!/bin/bash
# Runs ON a RunPod pod. Executes a stride-slice of the (deficit-end W, seed) grid, uploads each
# result JSON as it lands, then self-terminates. Same durability rule as the clone stage: the pod
# owns delivery, nothing depends on a Mac-side watcher surviving.
#
#   bash cp_sweep_bootstrap.sh <job_index> <n_jobs> <order> <total_steps>
#
# Grid is defined here so every pod agrees on it without extra plumbing. W=0 is the fresh
# baseline. Ends are log-spaced from below the 3-body acquisition point (100 steps) to 20x past
# the 33-body one (1000), which is the range the toy W-sweep says the effect turns on and peaks in.
set -x
JOBS=${1:?job indices, comma-separated}; N_JOBS=${2:?n jobs}; ORDER=${3:-9}; STEPS=${4:-100000}
WD=${5:-0.1}; WS=${6:-"0 250 500 1000 2000 5000 10000 20000"}; SEEDS=${7:-"0 1 2"}
POST=${8:-0}; CORPUS=${9:-}
TAG=$(echo "$JOBS-wd$WD-post$POST-c$CORPUS" | tr "," "-")   # one pod may cover several stride slices when GPU
                                   # capacity is scarce -- see the SUPPLY_CONSTRAINT note
case "$ORDER" in 3) NL=2;; 9) NL=4;; 33) NL=6;; *) echo "bad order"; exit 1;; esac
cd /root

apt-get update -qq && apt-get install -y -qq tmux >/dev/null 2>&1
pip install -q huggingface_hub numpy

python - "$NL" "$ORDER" "$CORPUS" <<'PY'
import os, sys, shutil, glob
from huggingface_hub import hf_hub_download, HfApi
nl, order = sys.argv[1], sys.argv[2]
os.makedirs("/root/results/nlp_cp", exist_ok=True)
for f in ["train.npy", "test.npy", f"factored{nl}.pt", "clone_factored2_sweep20.npy",
          "clone_factored4_sweep20.npy", "clone_factored6_sweep20.npy"]:
    p = hf_hub_download("bgradowhite/rende-nlp-fig1", f"results/{f}",
                        repo_type="dataset", local_dir="/root/dl")
    shutil.copy(p, f"/root/results/nlp_cp/{os.path.basename(f)}")
# every clone shard for this order
api = HfApi()
files = api.list_repo_files("bgradowhite/rende-nlp-cp", repo_type="dataset")
corpus = sys.argv[3] if len(sys.argv) > 3 else ""
shards = [f for f in files if f.startswith(f"clone{order}{corpus}_shard") and f.endswith(".npy")]
assert shards, f"no clone shards on HF for order {order}"
for f in shards:
    p = hf_hub_download("bgradowhite/rende-nlp-cp", f, repo_type="dataset",
                        local_dir="/root/dl2")
    shutil.copy(p, f"/root/results/nlp_cp/{os.path.basename(f)}")
print(f"data ready: {len(shards)} clone shards")
PY

nvidia-smi --query-gpu=name,ecc.errors.uncorrected.volatile.total --format=csv
python -c "import torch;assert torch.cuda.is_available();print('cuda OK',torch.cuda.get_device_name(0))" || {
  echo "GPU UNUSABLE - self-terminating"; runpodctl remove pod "$RUNPOD_POD_ID"; exit 1; }

i=0
for w in $WS; do
  for s in $SEEDS; do
    r=$((i % N_JOBS))
    case ",$JOBS," in *",$r,"*)
      echo "=== running W=$w seed=$s ==="
      PYTHONUNBUFFERED=1 python -u nlp_cp_sweep.py sweep --deficit-end "$w" --seed "$s" \
        --order "$ORDER" --total-steps "$STEPS" --wd "$WD" --post-steps "$POST" --corpus "$CORPUS" --outdir /root/results/nlp_cp \
        2>&1 | tee -a /root/job$TAG.log
      ;;
    esac
    i=$((i+1))
  done
done

python - "$TAG" <<'PY'
import sys
from huggingface_hub import HfApi
api = HfApi()
try:
    api.upload_file(path_or_fileobj=f"/root/job{sys.argv[1]}.log",
                    path_in_repo=f"logs/sweep_job{sys.argv[1]}.log",
                    repo_id="bgradowhite/rende-nlp-cp", repo_type="dataset")
except Exception as e:
    print("log upload failed", e)
PY

runpodctl remove pod "$RUNPOD_POD_ID"