Stop a neighbour's submission blocking an upload, and stop writing 150 GB of snapshots

tests/test_rf100vl_train.py

@@ -640,3 +640,42 @@ def test_keep_cache_opts_out_of_reclaiming_the_cache(tmp_path):
 
     assert cached.exists()
     assert freed == 2048
+
+
+def test_every_recipe_disables_periodic_snapshots():
+    """LibreYOLO's TrainConfig defaults to save_period=10, so families whose
+    trainer honours it write a full checkpoint every 10 epochs: ~1000 files and
+    150 GB across one campaign, on a 250 GB box. Nothing reads them. Resume
+    uses last.pt, selection uses best.pt, and the uploader ships best.pt. A
+    campaign already hit 97% disk on these alone."""
+    for family in set(list_families()):
+        recipe = rf100vl_train.load_recipe(
+            rf100vl_train.recipe_path_for_family(family),
+            family=family,
+        )
+        assert recipe["train"].get("save_period") == 0, (
+            f"{family} does not disable save_period; periodic snapshots will "
+            "fill the campaign box"
+        )
+
+
+def test_submission_recipe_sha_reads_only_this_model(tmp_path):
+    """Submissions land in one shared directory that accumulates every earlier
+    campaign. Reading the newest of all of them compared a neighbour's recipe
+    against this run and refused the upload."""
+    import json as _json
+    from va_bench import artifacts
+
+    def write(name, sha):
+        (tmp_path / name).write_text(
+            _json.dumps({"rf100vl": {"recipe_sha256": sha}}), encoding="utf-8"
+        )
+
+    mine = "a" * 64
+    theirs = "b" * 64
+    write("ec-s__pytorch__cuda__x__20260101T000000Z.json", mine)
+    # sorts after ours, and is what the old code would have picked
+    write("yolox-m__pytorch__cuda__x__20260909T000000Z.json", theirs)
+
+    assert artifacts._submission_recipe_sha(tmp_path, "ec-s") == mine
+    assert artifacts._submission_recipe_sha(tmp_path, "yolox-m") == theirs

va_bench/artifacts.py

@@ -381,11 +381,22 @@ class IncompletePublish(RuntimeError):
     """A publishable tier is missing artifacts that make the run reproducible."""
 
 
-def _submission_recipe_sha(submissions_dir: str | Path | None) -> str | None:
-    """recipe_sha256 recorded by the newest submission, or None."""
+def _submission_recipe_sha(
+    submissions_dir: str | Path | None,
+    model_key: str | None = None,
+) -> str | None:
+    """recipe_sha256 recorded by this model's newest submission, or None.
+
+    Submissions are written to one shared directory, so it accumulates every
+    earlier campaign's files. Reading the newest of ALL of them compared a
+    neighbouring model's recipe against this one and refused the upload: a
+    stale yolox submission blocked three ec-s attempts with a mismatch error
+    that named ec-s. Only this model's own submissions can describe its run.
+    """
     if not submissions_dir:
         return None
-    paths = sorted(Path(submissions_dir).glob("*.json"))
+    pattern = f"{model_key}__*.json" if model_key else "*.json"
+    paths = sorted(Path(submissions_dir).glob(pattern))
     for path in reversed(paths):
         try:
             with open(path, encoding="utf-8") as handle:
@@ -528,7 +539,7 @@ def add(path: Path, repo_path: str, *, required: bool = False) -> None:
     # The resolver falls back to the packaged family recipe when --recipe is
     # omitted, which is the wrong file for any run that used a custom one.
     if publishable and recipe_path and Path(recipe_path).is_file():
-        claimed = _submission_recipe_sha(submissions_dir)
+        claimed = _submission_recipe_sha(submissions_dir, model_key)
         if claimed:
             actual = _sha256(Path(recipe_path))
             if actual != claimed:

va_bench/recipes/rf100vl/deim.json

@@ -21,6 +21,7 @@
   },
   "sizes": {},
   "train": {
+    "save_period": 0,
     "imgsz": 640,
     "optimizer": "adamw",
     "lr0": 0.0001,

va_bench/recipes/rf100vl/deimv2.json

@@ -25,6 +25,7 @@
     "femto": {"imgsz": 416}
   },
   "train": {
+    "save_period": 0,
     "imgsz": 640,
     "optimizer": "adamw",
     "lr0": 0.0005,

va_bench/recipes/rf100vl/dfine.json

@@ -28,6 +28,7 @@
     "x": {"imgsz": 640}
   },
   "train": {
+    "save_period": 0,
     "optimizer": "adamw",
     "lr0": 0.0002,
     "weight_decay": 0.0001,

va_bench/recipes/rf100vl/ec.json

@@ -27,6 +27,7 @@
     "x": {"imgsz": 640, "lr0": 0.0005, "backbone_lr_mult": 0.01}
   },
   "train": {
+    "save_period": 0,
     "optimizer": "adamw",
     "weight_decay": 0.0001,
     "scheduler": "flat_cosine",

va_bench/recipes/rf100vl/picodet.json

@@ -26,6 +26,7 @@
     "l": {"imgsz": 640}
   },
   "train": {
+    "save_period": 0,
     "optimizer": "sgd",
     "lr0": 0.01,
     "momentum": 0.9,

va_bench/recipes/rf100vl/rfdetr.json

@@ -35,6 +35,7 @@
     }
   },
   "train": {
+    "save_period": 0,
     "optimizer": "adamw",
     "lr0": 0.0001,
     "weight_decay": 0.0001,

va_bench/recipes/rf100vl/rtdetr.json

@@ -21,6 +21,7 @@
   },
   "sizes": {},
   "train": {
+    "save_period": 0,
     "imgsz": 640,
     "optimizer": "adamw",
     "lr0": 0.0001,

va_bench/recipes/rf100vl/rtdetrv2.json

@@ -22,6 +22,7 @@
   },
   "sizes": {},
   "train": {
+    "save_period": 0,
     "imgsz": 640,
     "optimizer": "adamw",
     "lr0": 0.0001,

va_bench/recipes/rf100vl/rtdetrv4.json

@@ -21,6 +21,7 @@
   },
   "sizes": {},
   "train": {
+    "save_period": 0,
     "imgsz": 640,
     "optimizer": "adamw",
     "lr0": 0.0005,

va_bench/recipes/rf100vl/rtmdet.json

@@ -22,6 +22,7 @@
   },
   "sizes": {},
   "train": {
+    "save_period": 0,
     "imgsz": 640,
     "optimizer": "adamw",
     "lr0": 0.004,

va_bench/recipes/rf100vl/yolonas.json

@@ -26,6 +26,7 @@
     "l": {"imgsz": 640, "lr0": 0.0005}
   },
   "train": {
+    "save_period": 0,
     "optimizer": "adamw",
     "momentum": 0.9,
     "weight_decay": 0.00001,

va_bench/recipes/rf100vl/yolov7.json

@@ -24,6 +24,7 @@
     "b": {"imgsz": 640, "lr0": 0.01}
   },
   "train": {
+    "save_period": 0,
     "optimizer": "sgd",
     "momentum": 0.937,
     "weight_decay": 0.0005,

va_bench/recipes/rf100vl/yolov9-e2e.json

@@ -31,6 +31,7 @@
     "c": {"imgsz": 640, "lr0": 0.01}
   },
   "train": {
+    "save_period": 0,
     "optimizer": "sgd",
     "momentum": 0.937,
     "weight_decay": 0.0005,

va_bench/recipes/rf100vl/yolov9.json

@@ -28,6 +28,7 @@
     "c": {"imgsz": 640, "lr0": 0.01}
   },
   "train": {
+    "save_period": 0,
     "optimizer": "sgd",
     "momentum": 0.937,
     "weight_decay": 0.0005,

va_bench/recipes/rf100vl/yolox.json

@@ -29,6 +29,7 @@
     "x": {"imgsz": 640, "lr0": 0.01}
   },
   "train": {
+    "save_period": 0,
     "optimizer": "sgd",
     "momentum": 0.937,
     "weight_decay": 0.0005,