#!/usr/bin/env bash
# Headless CPU-pod worker: validate the mixed-Q2 artifact and pack it into GGUF.
# No SSH required: everything is driven by this script and evidence is pushed to Hugging Face.
#
# Inputs (environment):
#   HF_TOKEN        token with write access to EVIDENCE_REPO
#   EVIDENCE_REPO   default apetersson/v41-quant-worker (dataset)
#   ARTIFACT        default /workspace/v41-quant/q2-iq2-20260910
#   OUT_DIR         default /workspace/v41-quant/gguf-mixedq2
#   SOURCE_DIR      default /workspace/v41-download/models/DeepSeek-V4.1-Flash/2bc89ac599031fa673cab993f1df02fc4a98c673
#   LLAMA_PR_REF    default refs/pull/28696/head (V4.1 converter/gguf-py)
set -uo pipefail
trap 'echo "worker aborted rc=$? at line $LINENO"' ERR
STAMP=$(date -u +%Y%m%dT%H%M%SZ)
WORK=${WORK:-/workspace/v41-quant}
LOGDIR="$WORK/logs"
mkdir -p "$LOGDIR"
LOG="$LOGDIR/bootstrap-$STAMP.log"
exec > >(tee -a "$LOG") 2>&1
echo "=== v41 quant worker start $STAMP"

EVIDENCE_REPO=${EVIDENCE_REPO:-apetersson/v41-quant-worker}
ARTIFACT=${ARTIFACT:-$WORK/q2-iq2-20260910}
OUT_DIR=${OUT_DIR:-$WORK/gguf-mixedq2}
SOURCE_DIR=${SOURCE_DIR:-/workspace/v41-download/models/DeepSeek-V4.1-Flash/2bc89ac599031fa673cab993f1df02fc4a98c673}
LLAMA_PR_REF=${LLAMA_PR_REF:-refs/pull/28696/head}
LLAMA_DIR=/opt/llama.cpp
STATUS="$WORK/status.json"
PYTHONPATH_HINT="$LLAMA_DIR/gguf-py"
export PYTHONPATH="$PYTHONPATH_HINT"   # audit/mapping/pack all import gguf-py (PR #28696 head)

status() { # phase, detail
  python3 - "$STATUS" "$1" "$2" "$ARTIFACT" "$OUT_DIR" <<'PY'
import json, sys, time, os
path, phase, detail, artifact, out = sys.argv[1:6]
data = {}
if os.path.exists(path):
    data = json.load(open(path))
data.update(phase=phase, detail=detail, updated_at=time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime()),
            artifact=artifact, out_dir=out)
json.dump(data, open(path, "w"), indent=1)
PY
  echo "--- phase: $1 ($2)"
}

push_evidence() { # file...
  [ -n "${HF_TOKEN:-}" ] || { echo "no HF_TOKEN; skipping evidence push"; return 0; }
  for f in "$@"; do
    [ -f "$f" ] || continue
    python3 - "$f" "$EVIDENCE_REPO" "$STAMP" <<'PY'
import sys, os
from huggingface_hub import HfApi
path, repo, stamp = sys.argv[1:4]
api = HfApi(token=os.environ["HF_TOKEN"])
api.upload_file(path_or_fileobj=path, path_in_repo=f"runs/{stamp}/{os.path.basename(path)}",
                repo_id=repo, repo_type="dataset", commit_message=f"v41 quant worker: {os.path.basename(path)}")
print("pushed", os.path.basename(path))
PY
  done
}

status setup "installing toolchain"
export DEBIAN_FRONTEND=noninteractive
apt-get update -qq && apt-get install -y -qq --no-install-recommends \
  git cmake g++ make curl ca-certificates python3-pip procps >/tmp/apt.log 2>&1
python3 -m pip install --quiet --no-cache-dir numpy pyyaml tqdm huggingface_hub

status bundle "fetching worker bundle"
curl -fsSL --retry 6 --retry-delay 5 --retry-all-errors -H "Authorization: Bearer ${HF_TOKEN:-}" -o /tmp/v41-runtimes.tgz "https://huggingface.co/datasets/$EVIDENCE_REPO/resolve/main/v41-runtimes-20260910T1240.tgz" \
  && tar xzf /tmp/v41-runtimes.tgz -C "$WORK" --warning=no-unknown-keyword
SCRIPTS="$WORK/scripts"
ls "$SCRIPTS" >/dev/null || { echo "worker bundle missing $SCRIPTS"; exit 3; }

status llama "cloning llama.cpp $LLAMA_PR_REF"
rm -rf "$LLAMA_DIR"; mkdir -p "$LLAMA_DIR"
git -C "$LLAMA_DIR" init -q .
git -C "$LLAMA_DIR" remote add origin https://github.com/ggml-org/llama.cpp.git
git -C "$LLAMA_DIR" fetch -q --depth 1 origin "$LLAMA_PR_REF"
git -C "$LLAMA_DIR" checkout -q FETCH_HEAD
git -C "$LLAMA_DIR" log --oneline -1 | tee "$WORK/llama-ref.txt"

status ggml "building ggml (CPU)"
cmake -S "$LLAMA_DIR" -B "$LLAMA_DIR/build" -DCMAKE_BUILD_TYPE=Release \
  -DBUILD_SHARED_LIBS=ON -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_EXAMPLES=OFF \
  -DLLAMA_BUILD_TOOLS=OFF -DLLAMA_BUILD_SERVER=OFF >/tmp/cmake.log 2>&1 \
  && cmake --build "$LLAMA_DIR/build" --target ggml -j"$(nproc)" >/tmp/build.log 2>&1
echo "ggml build rc=$?"

status verify "verifying artifact digests"
python3 "$SCRIPTS/v41_quant_experts.py" verify --out "$ARTIFACT" --sample 0 | tee "$WORK/verify.txt"
VERIFY_RC=${PIPESTATUS[0]}

status audit "fidelity audit"
python3 "$SCRIPTS/v41_quant_audit.py" --source "$SOURCE_DIR" --out "$ARTIFACT" \
  --per-projection 8 --json "$WORK/audit.json" 2>&1 | tail -30 || echo "audit failed (non-fatal)"
push_evidence "$WORK/audit.json" "$WORK/verify.txt" "$WORK/llama-ref.txt"

status mapping "gguf mapping report"
PYTHONPATH="$PYTHONPATH_HINT" python3 "$SCRIPTS/v41_quant_to_gguf.py" --artifact "$ARTIFACT" \
  --config "$SOURCE_DIR/config.json" --check-mapping > "$WORK/gguf-pack-report.json" 2>&1
MAP_RC=$?
echo "mapping rc=$MAP_RC"; tail -25 "$WORK/gguf-pack-report.json"
push_evidence "$WORK/gguf-pack-report.json"

if [ "$MAP_RC" -ne 0 ]; then
  status failed "mapping incomplete"
  push_evidence "$LOG" "$STATUS"
  exit 4
fi

status pack "writing GGUF"
PYTHONPATH="$PYTHONPATH_HINT" TMPDIR="$WORK/tmp" python3 "$SCRIPTS/v41_quant_to_gguf.py" \
  --artifact "$ARTIFACT" --config "$SOURCE_DIR/config.json" --out-dir "$OUT_DIR" \
  --split-gb "${SPLIT_GB:-40}" | tee "$WORK/pack.txt"
PACK_RC=${PIPESTATUS[0]}

status sums "hashing shards"
( cd "$OUT_DIR" && sha256sum ./*.gguf > SHA256SUMS && wc -l SHA256SUMS ) | tee -a "$WORK/pack.txt"
python3 - "$OUT_DIR" "$WORK/gguf-summary.json" <<'PY'
import glob, json, os, sys
out, dest = sys.argv[1:3]
files = sorted(glob.glob(os.path.join(out, "*.gguf")))
record = {
    "files": [{"name": os.path.basename(f), "bytes": os.path.getsize(f)} for f in files],
    "total_bytes": sum(os.path.getsize(f) for f in files),
    "shards": len(files),
}
json.dump(record, open(dest, "w"), indent=1)
print(json.dumps({k: v for k, v in record.items() if k != "files"}, indent=1))
PY
push_evidence "$WORK/pack.txt" "$WORK/gguf-summary.json" "$OUT_DIR/gguf-validation.json" "$OUT_DIR/SHA256SUMS"

status "${PACK_RC:0:1}${VERIFY_RC:0:1}" "pack rc=$PACK_RC verify rc=$VERIFY_RC"
if [ "$PACK_RC" -eq 0 ]; then status done "gguf pack complete"; else status failed "pack rc=$PACK_RC"; fi
push_evidence "$LOG" "$STATUS"
echo "=== worker finished"