train: create the ckpt repo if absent before the first push

train/train_ounce100m.py

@@ -243,6 +243,10 @@ def main():
         ounce100m_credentials.install()
         from huggingface_hub import HfApi
         api = HfApi()
+        # Create-if-absent: an upload to a repo nobody made yet is a failure that lands 10 minutes into
+        # the first checkpoint, on a GPU session that is being billed while it waits.
+        api.create_repo(repo_id=args.hub_repo, repo_type=args.hub_repo_type, exist_ok=True,
+                        token=os.environ.get("HF_TOKEN"))
         say(f"checkpoint hub target {args.hub_repo} (prune={args.prune}, every {ckpt_every} steps)")
 
     tacc = TrainingArguments(