push_ckpts.sh

#!/bin/bash
# Incrementally push each COMPLETED checkpoint to HF as it saves — so we don't gamble on one giant
# 117GB upload at completion. Self-extracts HF_TOKEN from the box's onstart. Runs until training done.
eval "$(grep -oE 'export HF_HUB_ENABLE_HF_TRANSFER=1 HF_TOKEN=[^ ]+ WANDB_API_KEY=[^ ]+ WANDB_ENTITY=[^ ]+ XLA_PYTHON_CLIENT_MEM_FRACTION=[0-9.]+' /root/onstart.sh | head -1)"
export HF_HUB_DISABLE_XET=1
CK=/workspace/openpi/checkpoints/pi05_yam7h_ea/run_ax91
REPO=Kavin60606/fdstudio-run_ax91
PUSHED=/root/pushed_ckpts.txt; touch "$PUSHED"
echo "checkpoint pusher up @ $(date -u +%T) -> $REPO/checkpoints/"
while true; do
  for d in "$CK"/*/; do
    step=$(basename "$d")
    [[ "$step" =~ ^[0-9]+$ ]] || continue
    grep -qx "$step" "$PUSHED" && continue
    # skip a dir still being written (modified in the last 120s) — push only settled checkpoints
    [ -n "$(find "$d" -newermt '-120 seconds' 2>/dev/null | head -1)" ] && continue
    echo ">>> pushing step $step ($(du -sh "$d"|cut -f1)) @ $(date -u +%T)"
    if /.venv/bin/python -c "from huggingface_hub import HfApi; HfApi(token='$HF_TOKEN').upload_folder(folder_path='$d', repo_id='$REPO', repo_type='dataset', path_in_repo='checkpoints/$step')"; then
      echo "$step" >> "$PUSHED"; echo ">>> DONE step $step"
    else
      echo ">>> FAILED step $step (will retry next loop)"
    fi
  done
  if [ -f /workspace/.tdone ]; then echo "training finished — one final sweep then exit"; sleep 10
    for d in "$CK"/*/; do step=$(basename "$d"); [[ "$step" =~ ^[0-9]+$ ]] || continue; grep -qx "$step" "$PUSHED" && continue
      /.venv/bin/python -c "from huggingface_hub import HfApi; HfApi(token='$HF_TOKEN').upload_folder(folder_path='$d', repo_id='$REPO', repo_type='dataset', path_in_repo='checkpoints/$step')" && echo "$step" >> "$PUSHED" && echo "final-swept $step"; done
    echo "ALL PUSHED: $(sort -n $PUSHED | tr '\n' ' ')"; break
  fi
  sleep 90
done