Add skippable success

.gitignore

@@ -0,0 +1 @@
+outputs/

README.md

@@ -42,6 +42,38 @@ apex-test-suite/
 
 ### 1) Download the dataset from HuggingFace
 
+### Hugging Face auth (recommended for private/gated repos)
+
+If you see `401/403 Unauthorized` from Hugging Face (dataset downloads or model/component downloads), export a token in your shell **before** running the suite/API.
+
+- **Linux / macOS (bash/zsh)**
+
+```bash
+export HF_TOKEN="hf_xxx"
+# or (also supported)
+export HUGGING_FACE_HUB_TOKEN="hf_xxx"
+```
+
+- **Windows (PowerShell)**
+
+```powershell
+$env:HF_TOKEN="hf_xxx"
+# or
+$env:HUGGING_FACE_HUB_TOKEN="hf_xxx"
+```
+
+- **Windows (cmd.exe)**
+
+```bat
+set HF_TOKEN=hf_xxx
+REM or
+set HUGGING_FACE_HUB_TOKEN=hf_xxx
+```
+
+Notes:
+- `hf auth login` stores a token on disk, but exporting `HF_TOKEN` is the most reliable option for test runs (especially when subprocesses/Ray workers are involved).
+- For **gated models**, you must also request/accept access on the model page in Hugging Face; a valid token alone can still return 401/403.
+
 Pick one:
 
 - **Option A: `huggingface_hub` snapshot download**
@@ -130,6 +162,17 @@ Run a subset by filename substring:
 .venv/bin/python apps/api/test_suite/run_suite.py --filter seedvr2
 ```
 
+Resume / rerun only failures (skip successes):
+
+This treats a test as “successful” if its expected artifact already exists under `outputs/`
+(e.g. `outputs/<json_stem>.png` for `image/` tests, `outputs/<json_stem>.mp4` for `video/` / `upscalers/`).
+
+```bash
+.venv/bin/python apps/api/test_suite/run_suite.py --skip-successes
+# (alias)
+.venv/bin/python apps/api/test_suite/run_suite.py --only-failed
+```
+
 Run a single test JSON by absolute path:
 
 ```bash

run_suite.py

@@ -152,6 +152,23 @@ def _match_filters(p: Path, kind: Optional[str], name_contains: Optional[str]) -
     return True
 
 
+def _expected_output_path_for_test(test_json: Path, outputs_dir: Path) -> Path:
+    """
+    Determine the expected output artifact path for a test JSON based on its kind.
+
+    Canonical naming matches `run_one.py` behavior:
+      outputs/<json_stem>.<ext>
+
+    Default mapping:
+    - image/*.json     -> .png
+    - video/*.json     -> .mp4
+    - upscalers/*.json -> .mp4
+    """
+    kind = test_json.parts[-2] if len(test_json.parts) >= 2 else ""
+    ext = ".png" if kind == "image" else ".mp4"
+    return outputs_dir / f"{test_json.stem}{ext}"
+
+
 def _parse_child_json(stdout: str) -> Dict[str, Any]:
     # Child prints one JSON object as the last line. Be defensive in case logs leaked.
     lines = [ln.strip() for ln in stdout.splitlines() if ln.strip()]
@@ -253,6 +270,23 @@ def main() -> int:
             "'on' enables previews for all tests."
         ),
     )
+
+    parser.add_argument(
+        "--skip-count",
+        type=int,
+        default=0,
+        help="Number of tests to skip at the beginning.",
+    )
+    parser.add_argument(
+        "--only-failed",
+        "--skip-successes",
+        dest="only_failed",
+        action="store_true",
+        help=(
+            "Skip tests that already have their expected output artifact in outputs/. "
+            "This effectively runs only tests that are missing outputs (i.e. previously failed/incomplete)."
+        ),
+    )
     args = parser.parse_args()
 
     api_root = Path(__file__).resolve().parents[1]
@@ -292,10 +326,22 @@ def main() -> int:
         if not nunchaku_available:
             tests = [p for p in tests if "nunchaku" not in p.name.lower()]
 
+    # Optionally filter to tests missing their expected output artifact in outputs/.
+    if args.only_failed and not args.json:
+        before = len(tests)
+        tests = [
+            p
+            for p in tests
+            if not _expected_output_path_for_test(p, outputs_dir).exists()
+        ]
+        print(
+            f"Skip-successes mode: selected {len(tests)} tests missing outputs (from {before} discovered)."
+        )
+
     started = time.perf_counter()
     results: List[Dict[str, Any]] = []
     
-    for idx, p in enumerate(tests):
+    for idx, p in enumerate(tests[args.skip_count:]):
         env = os.environ.copy()
         env["PYTHONUNBUFFERED"] = "1"
         # Reduce Ray log de-duping so you can see what's happening during long runs.
@@ -357,6 +403,7 @@ def main() -> int:
         "skipped": len(skipped),
         "total_elapsed_s": total_s,
         "arch": (results[0].get("arch") if results else None),
+        "only_failed": bool(args.only_failed and not args.json),
         "results": results,
     }