Upload .mirror/mirror_worker.py with huggingface_hub

.mirror/mirror_worker.py

@@ -159,9 +159,13 @@ def main() -> None:
         if number % WORKER_COUNT == WORKER_INDEX
         and (item.get("Type") or item.get("type")) != "tree"
     ]
-    # Each shard is processed in deterministic manifest order.  Resume at its
-    # first absent path instead of walking tens of thousands of committed files
-    # after every hourly job restart.
+    # Prioritize payload throughput. ModelScope's path order begins with a
+    # massive collection of sub-kilobyte metrics JSONs; sending those first
+    # strands the workers in metadata while multi-gigabyte payloads wait.
+    # The order remains deterministic, and uploaded paths are still skipped.
+    selected.sort(key=lambda item: int(item.get("Size") or item.get("size") or 0), reverse=True)
+    # Resume at the first not-yet-uploaded file instead of walking committed
+    # files after every hourly job restart.
     shard_total = len(selected)
     resume_at = next(
         (