Update main.py

main.py

@@ -16,28 +16,53 @@ import torch
 import numpy as np
 import faiss
 
-# ── Compatibility patch ────────────────────────────────────────────────────────
-# `find_pruneable_heads_and_indices` was removed from transformers.pytorch_utils
-# in newer releases, but the jina-bert-v2 custom modeling code still imports it
-# from there.  We re-inject the original implementation before anything loads.
+# ── Compatibility patches ──────────────────────────────────────────────────────
+# The jina-bert-v2 custom modeling code was written against an older transformers
+# API. Two things were removed / tightened in newer releases:
+#
+# 1. `find_pruneable_heads_and_indices` was removed from transformers.pytorch_utils.
+# 2. `PretrainedConfig` no longer sets is_decoder / add_cross_attention as instance
+#    defaults in __init__. A tightened __getattribute__ now raises AttributeError
+#    instead of the old silent fallback, breaking JinaBertConfig access patterns.
+#
+# Both patches are guarded with hasattr/flag checks so they are no-ops if a future
+# transformers version re-adds these symbols.
+
 import transformers.pytorch_utils as _pt_utils
 if not hasattr(_pt_utils, "find_pruneable_heads_and_indices"):
-    from typing import Set, List, Tuple
     def _find_pruneable_heads_and_indices(
-        heads: List[int],
-        n_heads: int,
-        head_size: int,
-        already_pruned_heads: Set[int],
-    ) -> Tuple[Set[int], torch.LongTensor]:
+        heads, n_heads: int, head_size: int, already_pruned_heads
+    ):
+        import torch
         mask = torch.ones(n_heads, head_size)
         heads = set(heads) - already_pruned_heads
         for head in heads:
             head = head - sum(1 if h < head else 0 for h in already_pruned_heads)
             mask[head] = 0
         mask = mask.view(-1).contiguous().eq(1)
-        index: torch.LongTensor = torch.arange(len(mask))[mask].long()
+        index = torch.arange(len(mask))[mask].long()
         return heads, index
     _pt_utils.find_pruneable_heads_and_indices = _find_pruneable_heads_and_indices
+
+import transformers.configuration_utils as _cfg_utils
+_PC = _cfg_utils.PretrainedConfig
+if not hasattr(_PC, "_jina_compat_patched"):
+    # Attributes that used to be set in PretrainedConfig.__init__ with defaults
+    # but were removed from the base class in newer transformers versions.
+    _LEGACY_DEFAULTS = {
+        "is_decoder":               False,
+        "add_cross_attention":      False,
+        "cross_attention_hidden_size": None,
+        "use_cache":                True,
+    }
+    def _pc_getattr(self, key: str):
+        if key in _LEGACY_DEFAULTS:
+            return _LEGACY_DEFAULTS[key]
+        raise AttributeError(
+            f"'{type(self).__name__}' object has no attribute '{key}'"
+        )
+    _PC.__getattr__ = _pc_getattr
+    _PC._jina_compat_patched = True
 # ──────────────────────────────────────────────────────────────────────────────
 
 from fastapi import FastAPI, HTTPException, UploadFile, File, Form