Fix DynamicCacheWithRepeat AttributeError on transformers Cache refactor

minference/modules/kvcompression.py

@@ -370,6 +370,11 @@ def get_kv_cluster_class_config(self, layer_idx: int, head_idx: int = None):
 class DynamicCacheWithRepeat(DynamicCache):
     def __init__(self, config, *args, **kwargs):
         super().__init__(*args, **kwargs)
+        # update()/get_seq_length() read these directly and never touch DynamicCache's own
+        # storage, so self-init them like BaseKVCache above instead of relying on the parent.
+        self._seen_tokens = 0
+        self.key_cache: List[torch.Tensor] = []
+        self.value_cache: List[torch.Tensor] = []
         self.temp_key_cache = []
         self.temp_value_cache = []
 

tests/test_dynamic_cache_with_repeat.py

@@ -0,0 +1,71 @@
+# Copyright (c) 2024-2025 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+
+import importlib
+import importlib.machinery
+import os
+import sys
+import types
+import unittest
+
+import torch
+
+# `import minference` needs vllm, and kivi.py/retr_attn.py need kivi_gemv/papyfaiss
+# (guarded by a check that misfires truthy on current transformers), none of which
+# DynamicCacheWithRepeat touches. Load the module under test directly instead.
+for _name in ("kivi_gemv", "papyfaiss"):
+    if _name not in sys.modules:
+        _stub = types.ModuleType(_name)
+        _stub.__spec__ = importlib.machinery.ModuleSpec(_name, loader=None)
+        sys.modules[_name] = _stub
+
+_minference_dir = os.path.join(
+    os.path.dirname(os.path.dirname(os.path.abspath(__file__))), "minference"
+)
+if "minference" not in sys.modules:
+    _pkg = types.ModuleType("minference")
+    _pkg.__path__ = [_minference_dir]
+    sys.modules["minference"] = _pkg
+if "minference.modules" not in sys.modules:
+    _modules_pkg = types.ModuleType("minference.modules")
+    _modules_pkg.__path__ = [os.path.join(_minference_dir, "modules")]
+    sys.modules["minference.modules"] = _modules_pkg
+
+DynamicCacheWithRepeat = importlib.import_module(
+    "minference.modules.kvcompression"
+).DynamicCacheWithRepeat
+
+
+class DynamicCacheWithRepeatTest(unittest.TestCase):
+    def test_empty_cache_seq_length_is_zero(self):
+        cache = DynamicCacheWithRepeat(config=None)
+        self.assertEqual(cache.get_seq_length(), 0)
+
+    def test_prefill_then_decode_updates_seq_length(self):
+        cache = DynamicCacheWithRepeat(config=None)
+        key_states = torch.randn(1, 2, 5, 4)
+        value_states = torch.randn(1, 2, 5, 4)
+        out_k, out_v = cache.update(
+            key_states,
+            value_states,
+            layer_idx=0,
+            cache_kwargs={"update_global_past_kv": True},
+        )
+        torch.testing.assert_close(out_k, key_states)
+        torch.testing.assert_close(out_v, value_states)
+        self.assertEqual(cache.get_seq_length(), 5)
+
+        # decoding step: one more token appended to layer 0
+        key_states2 = torch.randn(1, 2, 1, 4)
+        value_states2 = torch.randn(1, 2, 1, 4)
+        cache.update(
+            key_states2,
+            value_states2,
+            layer_idx=0,
+            cache_kwargs={"update_global_past_kv": True},
+        )
+        self.assertEqual(cache.get_seq_length(), 6)
+
+
+if __name__ == "__main__":
+    unittest.main()