train: hubckpt -- hubckpt.ensure_repo writes LFS attributes before the first checkpoint push (found by P1)

train/hubckpt.py

@@ -117,6 +117,39 @@ def verify_checkpoint(repo, dirpath, path_in_repo, token=None, repo_type="datase
             "mismatch": mismatch[:10], "extra": extra[:10], "readback": rb}
 
 
+def ensure_repo(repo, api, token=None, repo_type="dataset"):
+    """Create the checkpoint repo and give it LFS patterns, in that order, before anything is pushed.
+
+    A 1.6 GB checkpoint is far over the 10 MB non-LFS limit, and a fresh repo has no .gitattributes, so a
+    first push fails with "should be tracked by LFS" -- ten minutes into the first checkpoint of a GPU
+    session that is being billed while it waits. Found by preflight P1 on CPU, which is the whole point
+    of running the cycle with mock bytes before the real run.
+    """
+    created = False
+    try:
+        api.create_repo(repo_id=repo, repo_type=repo_type, exist_ok=True, token=token)
+        created = True
+    except Exception as e:
+        # exist_ok=True should absorb the common case; anything else is worth seeing, not swallowing.
+        print(f"ensure_repo create: {type(e).__name__}: {str(e)[:160]}", flush=True)
+    have = set()
+    try:
+        have = set(api.list_repo_files(repo_id=repo, repo_type=repo_type))
+    except Exception as e:
+        print(f"ensure_repo listing failed ({type(e).__name__}); writing attributes anyway", flush=True)
+    if ".gitattributes" not in have:
+        body = ("*.safetensors filter=lfs diff=lfs merge=lfs -text\n"
+                "*.bin filter=lfs diff=lfs merge=lfs -text\n"
+                "*.pt filter=lfs diff=lfs merge=lfs -text\n"
+                "*.npy filter=lfs diff=lfs merge=lfs -text\n"
+                "*.ckpt filter=lfs diff=lfs merge=lfs -text\n")
+        api.upload_file(path_or_fileobj=body.encode(), path_in_repo=".gitattributes", repo_id=repo,
+                        repo_type=repo_type, commit_message="track checkpoint tensors with LFS",
+                        token=token)
+    return {"repo": repo, "created": created, "had_gitattributes": ".gitattributes" in have,
+            "files_before": len(have)}
+
+
 def push_and_prune(repo, dirpath, path_in_repo, api, repo_type="dataset", token=None,
                    prune=True, min_free_gb=6.0):
     """One full §3.13 cycle. Raises if verification fails and pruning was requested -- the local copy is