pin: train/hubckpt.py (RepoMissing resume branch, disk gate, PHASE4_PREP_ONLY)

train/hubckpt.py

@@ -52,6 +52,12 @@ def local_manifest(dirpath):
     return out
 
 
+class RepoMissing(RuntimeError):
+    """The Hub says this repository does not exist -- distinct from "the listing failed", which is what
+    RuntimeError means here. Callers that treat a failure as an empty repo would restart training at step
+    0 (E-018), so the two must stay distinguishable at the except clause."""
+
+
 def hub_listing(repo, repo_type="dataset", revision="main", token=None, page=None):
     """{path_in_repo: (bytes, sha256_or_None)} straight from the Hub's tree endpoint, ALL pages.
 
@@ -71,6 +77,13 @@ def hub_listing(repo, repo_type="dataset", revision="main", token=None, page=Non
         with urllib.request.urlopen(req, timeout=120) as r:
             entries = json.loads(r.read().decode())
             link = r.headers.get("Link") or ""
+    except urllib.error.HTTPError as e:
+        # 404 here means the repository itself has never been created, which is a different fact from a
+        # failed lookup -- and the first Phase 4 session depends on telling them apart.
+        if e.code == 404:
+            raise RepoMissing(f"hub repo {repo} does not exist (HTTP 404)") from e
+        raise RuntimeError(f"hub listing failed for {repo}@{revision}: HTTPError {e.code}: "
+                           f"{str(e)[:160]}") from e
     except Exception as e:
         raise RuntimeError(f"hub listing failed for {repo}@{revision}: {type(e).__name__}: "
                            f"{str(e)[:160]}") from e