Merge pull request #8 from PyTs1n9/fix/issue-6

Accept export UI, cache checks, decord compatibility and previously integrated SQLite/UUID fixes. Preserve port and GPU behavior; exclude local database; fix cross-platform launcher. Release v1.5.16.

Author: Somnusochi

CHANGELOG.md

@@ -1,5 +1,14 @@
 # Changelog
 
+## v1.5.16 (2026-09-09)
+
+- Merge: PR #8 by PyTs1n9 with maintainer compatibility adjustments.
+- Feat: add full-history YOLO Seg export to the training panel.
+- Fix: register the bundled decord fallback on macOS/Windows.
+- Fix: check required model cache files before skipping downloads.
+- Fix: resolve frontend launch commands across Windows and Unix platforms.
+- Retain existing port and GPU behavior; exclude the contributor's local database.
+
 ## v1.5.15 (2026-09-09)
 
 - Fix: normalize repository string IDs to UUIDs for SQLite reads and annotation writes (PyTs1n9, PR #8).

backend/app/services/locate_anything.py

@@ -29,9 +29,15 @@ def _ensure_decord_stub():
     if _decord_loaded:
         return
     if sys.platform in ("darwin", "win32"):
-        stub_dir = str(Path(settings.resolved_model_dir) / "stubs")
-        if Path(stub_dir).exists() and stub_dir not in sys.path:
-            sys.path.insert(0, stub_dir)
+        try:
+            import decord  # type: ignore[import-not-found]  # noqa: F401
+        except ImportError:
+            import decord_stub
+
+            # Transformers checks optional imports before loading remote model
+            # code. Register the bundled image-only compatibility shim under
+            # the package name it expects.
+            sys.modules["decord"] = decord_stub
     _decord_loaded = True
 
 
@@ -393,6 +399,7 @@ def detect(image_path: str | Path, categories: list[str]) -> dict:
         try:
             result = worker.detect(img, categories)
             raw_text = result["answer"]
+            logger.debug("Raw model response: %r", raw_text)
         except Exception as exc:
             raise InferenceError(f"Model inference failed: {exc}") from exc
 

backend/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "locate-anything-backend"
-version = "1.5.15"
+version = "1.5.16"
 description = "Backend API for VLM-AutoYOLO - VLM-powered auto-labeling and YOLO training platform"
 requires-python = ">=3.12"
 dependencies = [

backend/tests/test_cli.py

@@ -16,6 +16,38 @@
 SPEC.loader.exec_module(cli)
 
 
+def test_partial_model_cache_downloads(monkeypatch, tmp_path):
+    model = tmp_path / "model"
+    model.mkdir()
+    (model / "config.json").write_text("{}")
+    calls = []
+    monkeypatch.setattr(cli, "BACKEND", tmp_path)
+    monkeypatch.setattr(cli, "run", lambda *a, **k: calls.append(a))
+    cli.download_vlm()
+    assert calls
+
+
+@pytest.mark.parametrize("windows", [True, False])
+def test_frontend_launch_resolves_tool(monkeypatch, windows):
+    import urllib.request
+    from types import SimpleNamespace
+
+    calls = []
+    monkeypatch.setattr(cli, "IS_WIN", windows)
+    monkeypatch.setattr(cli, "is_port_open", lambda _: False)
+    monkeypatch.setattr(cli.time, "sleep", lambda _: None)
+    monkeypatch.setattr(cli.shutil, "which", lambda _: "pnpm.cmd" if windows else "/bin/pnpm")
+    monkeypatch.setattr(urllib.request, "urlopen", lambda *a, **k: None)
+    monkeypatch.setattr(
+        cli.subprocess,
+        "Popen",
+        lambda cmd, **k: calls.append(cmd) or SimpleNamespace(poll=lambda: None),
+    )
+    cli.start_frontend()
+    expected = ["cmd", "/d", "/c", "pnpm.cmd"] if windows else ["/bin/pnpm"]
+    assert calls[0] == [*expected, "dev", "--port", str(cli.FRONTEND_PORT)]
+
+
 def test_failed_migration_stops_setup(monkeypatch, capsys):
     monkeypatch.setattr(cli, "ALEMBIC", sys.executable)
     with pytest.raises(subprocess.CalledProcessError):

backend/tests/test_locate_anything.py

@@ -5,6 +5,27 @@
 from app.services.locate_anything import parse_boxes
 
 
+def test_bundled_decord_fallback(monkeypatch):
+    import builtins
+    import sys
+
+    from app.services import locate_anything
+
+    original = builtins.__import__
+
+    def without_decord(name, *args, **kwargs):
+        if name == "decord":
+            raise ImportError("decord unavailable")
+        return original(name, *args, **kwargs)
+
+    monkeypatch.setattr(builtins, "__import__", without_decord)
+    monkeypatch.setattr(sys, "platform", "win32")
+    monkeypatch.delitem(sys.modules, "decord", raising=False)
+    monkeypatch.setattr(locate_anything, "_decord_loaded", False)
+    locate_anything._ensure_decord_stub()
+    assert sys.modules["decord"].VideoReader
+
+
 def test_parse_boxes_single():
     raw = "<ref>cat</ref><box><100><200><300><400></box>"
     boxes = parse_boxes(raw, 1000, 1000)

cli.py

@@ -214,10 +214,29 @@ def check_sam3_token():
 def download_vlm():
     step("VLM model (LocateAnything-3B, ~6GB)...")
     model_dir = BACKEND / "model"
-    if model_dir.exists() and any(model_dir.iterdir()):
+    required_files = (
+        "config.json",
+        "model.safetensors.index.json",
+        "model-00001-of-00002.safetensors",
+        "model-00002-of-00002.safetensors",
+        "modeling_locateanything.py",
+        "modeling_qwen2.py",
+        "modeling_vit.py",
+        "preprocessor_config.json",
+        "processing_locateanything.py",
+        "processor_config.json",
+        "special_tokens_map.json",
+        "tokenizer_config.json",
+        "vocab.json",
+    )
+    if (
+        model_dir.exists()
+        and not any(model_dir.glob("*.segments.json"))
+        and all((model_dir / name).is_file() for name in required_files)
+    ):
         ok("already cached")
         return
-    print(f"  Downloading to {model_dir}... (~10-30 min)")
+    print(f"  Downloading missing model files to {model_dir}... (~10-30 min)")
     try:
         run(
             [str(PYTHON), "-c",
@@ -316,7 +335,7 @@ def start_frontend():
         warn(f"Port {FRONTEND_PORT} already in use. Try: FRONTEND_PORT={FRONTEND_PORT + 1} python3 cli.py start")
         return None
     proc = subprocess.Popen(
-        ["pnpm", "dev", "--port", str(FRONTEND_PORT)],
+        _tool_command("pnpm", "dev", "--port", str(FRONTEND_PORT)),
         cwd=FRONTEND,
         env={**os.environ, "VITE_BACKEND_PORT": str(BACKEND_PORT)},
         stdout=subprocess.DEVNULL,

docs/2026-09-09-pr8-integration-change.md

@@ -0,0 +1,22 @@
+# PR #8 integration review
+
+Accept the full-history export UI, bundled decord fallback and model cache
+validation, alongside the already accepted SQLite/UUID fixes. Retain existing
+export API defaults. Resolve frontend launching through the cross-platform
+helper rather than hardcoding pnpm.cmd. Keep port 8000 and exclude local DB.
+
+Do not apply automatic CPU offload or reduce generation length/image size:
+the custom model generation loop directly calls vision/MLP modules and needs
+hardware validation for dispatch; reducing tokens caps the detection output
+for all devices. These are separate tuning changes without accuracy evidence.
+
+Acceptance: export button calls full-history API, bundled stub imports when
+decord is absent, incomplete caches trigger download, frontend launcher works
+on both platforms. Preserve original PR ancestry in the merge commit.
+
+RED: four backend failures (cache, two launcher platforms, decord) and missing
+frontend export button. GREEN: 83 backend tests passed, 20 skipped; 43 frontend
+tests passed. Backend lint/format and frontend lint/TypeScript/build passed.
+REFACTOR: import ordering and test formatting. GPU inference on physical CUDA
+was not verified, so GPU tuning is excluded. No persisted schema change beyond previously
+accepted compatibility fixes. Rollback via reverting merge; do not overwrite DB.

frontend/package.json

@@ -1,7 +1,7 @@
 {
   "name": "frontend",
   "private": true,
-  "version": "1.5.15",
+  "version": "1.5.16",
   "type": "module",
   "scripts": {
     "dev": "vite",

frontend/src/auto-imports.d.ts

@@ -85,6 +85,7 @@ declare global {
   const downloadModelUrl: typeof import('./services/api').downloadModelUrl
   const downloadOnnxUrl: typeof import('./services/api').downloadOnnxUrl
   const downloadYoloTxt: typeof import('./lib/yoloExport').downloadYoloTxt
+  const exportAll: typeof import('./services/api').exportAll
   const exportBatch: typeof import('./services/api').exportBatch
   const exportSingleUrl: typeof import('./services/api').exportSingleUrl
   const extractKeyframes: typeof import('./services/api').extractKeyframes

frontend/src/components/TrainingPanel/index.test.tsx

@@ -10,6 +10,11 @@ vi.mock("@/store/useAppStore", () => ({
 }));
 
 describe("TrainingPanel", () => {
+  it("offers full-history export when records exist outside the loaded page", () => {
+    const { getByRole } = render(<TrainingPanel detections={[]} total={100}
+      hasNextPage={true} isFetchingNextPage={false} fetchNextPage={() => {}} />);
+    expect(getByRole("button", { name: /downloadAllYoloSeg|Export All YOLO Seg|导出全部/ })).toBeTruthy();
+  });
   it("renders without crashing", () => {
     // Note: Provide basic props if needed, or mock stores/hooks
     // This is a basic boilerplate test