train: hubckpt -- readback counts as verification; GQA kv must divide heads; pargs reports before optional probe

train/hubckpt.py

@@ -96,25 +96,30 @@ def verify_checkpoint(repo, dirpath, path_in_repo, token=None, repo_type="datase
         if hsz != lsz or (comparable and hsha != lsha):
             mismatch.append({"path": k, "local": [lsz, lsha[:16]], "hub": [hsz, (hsha or "")[:16]]})
     rb = None
+    rb_ok = False
     if readback_one and have:
-        # Download the smallest file back over an anonymous connection: cheapest possible proof that the
-        # bytes are readable, not merely listed.
+        # Download one file back over an anonymous connection: cheapest possible proof that the bytes are
+        # *readable*, not merely listed (§3.13 is explicit that listing is not verification).
         pick = min(have, key=lambda k: have[k][0])
         tmp = os.path.join("/tmp", "hubverify_" + os.path.basename(pick))
-        url = f"{HUB}/{repo_type}s/{repo}/resolve/{prefix}{pick}"
+        url = f"{HUB}/{repo_type}s/{repo}/resolve/main/{prefix}{pick}"
         try:
             urllib.request.urlretrieve(url, tmp)
             got = sha256_file(tmp)
-            rb = {"path": pick, "bytes": os.path.getsize(tmp),
-                  "sha_matches_local": got == want[pick][1] if pick in want else None}
+            rb_ok = bool(pick in want and got == want[pick][1])
+            rb = {"path": pick, "bytes": os.path.getsize(tmp), "sha_matches_manifest": rb_ok}
         except Exception as e:
-            rb = {"path": pick, "error": f"{type(e).__name__}: {str(e)[:160]}"}
+            rb = {"path": pick, "url": url,
+                  "error": f"{type(e).__name__}: {str(e)[:160]}"}
         finally:
             if os.path.exists(tmp):
                 os.remove(tmp)
-    ok = bool(want) and not missing and not mismatch
+    # ok must include the read-back: a repo whose files are listed but not downloadable is not a
+    # checkpoint, and pruning into that state loses the run.
+    ok = bool(want) and not missing and not mismatch and (rb_ok or not readback_one)
     return {"ok": ok, "n_local": len(want), "n_hub": len(have), "missing": missing[:10],
-            "mismatch": mismatch[:10], "extra": extra[:10], "readback": rb}
+            "mismatch": mismatch[:10], "extra": extra[:10], "readback": rb,
+            "readback_ok": rb_ok}
 
 
 def ensure_repo(repo, api, token=None, repo_type="dataset"):