setup.sh: one-shot bootstrap (workspace sync + hf_transfer fast downloads)

setup.sh

@@ -35,14 +35,19 @@ PYTHON_BIN="${VSI_PYTHON_BIN:-python3.11}"
 
 SKIP_MODELS=0
 SKIP_DATA=0
+SKIP_WORKSPACE=0
+WITH_CACHES=0
 FORCE=0
 ASSUME_YES=0
 HF_TOKEN="${HF_TOKEN:-}"
+WORKSPACE_REPO="${VSI_WORKSPACE_REPO:-AntonioJun/workspace}"
 
 for arg in "$@"; do
   case "$arg" in
     --skip-models) SKIP_MODELS=1 ;;
     --skip-data) SKIP_DATA=1 ;;
+    --skip-workspace) SKIP_WORKSPACE=1 ;;
+    --with-caches) WITH_CACHES=1 ;;
     --force) FORCE=1 ;;
     -y|--yes) ASSUME_YES=1 ;;
     --token=*) HF_TOKEN="${arg#--token=}" ;;
@@ -271,6 +276,15 @@ fi
 # `hf`; prefer `hf` wherever it exists and only fall back to the legacy name on
 # old installs. A silent legacy-CLI failure here previously left VSI-Bench and
 # every checkpoint undownloaded while the rest of setup continued.
+# Fast multi-threaded downloads: hf_transfer accelerates every hf download below
+# (models and datasets alike) by an order of magnitude on fat pipes.
+PIP_BIN="$VENV_ROOT/bin/pip"
+[ -x "$PIP_BIN" ] || PIP_BIN="${VLM_VENV_ROOT}/bin/pip"
+if [ -x "$PIP_BIN" ]; then
+  "$PIP_BIN" install -q hf_transfer huggingface_hub >/dev/null 2>&1 || warn "hf_transfer install failed; downloads will be slower"
+fi
+export HF_HUB_ENABLE_HF_TRANSFER=1
+
 HF_CLI=""
 for candidate in "$VENV_ROOT/bin/hf" "${PERCEPTION_VENV_ROOT}/bin/hf" \
     "$VENV_ROOT/bin/huggingface-cli" "${PERCEPTION_VENV_ROOT}/bin/huggingface-cli"; do
@@ -301,6 +315,58 @@ hf_download() {
   fi
 }
 
+# ---------------------------------------------------------------------------
+# 6a. Workspace sync -- pull this project's own code + spatial codes from the
+#     backup repo ($WORKSPACE_REPO), so a fresh pod needs ONLY this script to
+#     bootstrap. Uses per-folder tree listing (never enumerates the whole repo,
+#     which is huge) and skips files already present. --skip-workspace disables;
+#     --with-caches additionally pulls the SAM3/DA3 raw caches (large; only
+#     needed to run inference/encoder, not the VLM harnesses).
+# ---------------------------------------------------------------------------
+if [ "$SKIP_WORKSPACE" -eq 0 ]; then
+  log "Syncing workspace code + data from $WORKSPACE_REPO"
+  PY_SYNC="$VENV_ROOT/bin/python"
+  [ -x "$PY_SYNC" ] || PY_SYNC="${VLM_VENV_ROOT}/bin/python"
+  WITH_CACHES="$WITH_CACHES" WORKSPACE_REPO="$WORKSPACE_REPO" \
+    WORKSPACE_ROOT="$WORKSPACE_ROOT" "$PY_SYNC" - <<'PYSYNC' || die "workspace sync failed"
+import os
+from pathlib import Path
+from huggingface_hub import HfApi, hf_hub_download
+
+repo = os.environ["WORKSPACE_REPO"]
+root = Path(os.environ["WORKSPACE_ROOT"])
+token = os.environ.get("HF_TOKEN") or None
+api = HfApi(token=token)
+folders = [
+    "harness", "symbolic", "analysis", "corruption", "calibration",
+    "encoder", "inference", "tests", "data/spatial codes",
+]
+if os.environ.get("WITH_CACHES") == "1":
+    folders.append("data/caches")
+top_files = ["README.md", "backup.py", "setup.sh"]
+
+def fetch(rel):
+    dest = root / rel
+    if dest.is_file() and dest.stat().st_size > 0:
+        return 0
+    hf_hub_download(repo, rel, repo_type="dataset", local_dir=str(root), token=token)
+    return 1
+
+for folder in folders:
+    files = [
+        e.path for e in api.list_repo_tree(repo, repo_type="dataset",
+                                           path_in_repo=folder, recursive=True)
+        if e.__class__.__name__ == "RepoFile"
+    ]
+    got = sum(fetch(rel) for rel in files)
+    print(f"[{folder}] {len(files)} files ({got} downloaded, rest already present)", flush=True)
+for rel in top_files:
+    fetch(rel)
+print("workspace sync complete", flush=True)
+PYSYNC
+  chmod +x "$WORKSPACE_ROOT/setup.sh" "$WORKSPACE_ROOT/backup.py" 2>/dev/null || true
+fi
+
 if [ "$SKIP_DATA" -eq 0 ]; then
   hf_download "nyu-visionx/VSI-Bench" dataset "$DATA_ROOT/VSI-Bench"