dirac/evals/cline/cline_refactor_DynamicCache at master · dirac-run/dirac · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
diff --git a/src/transformers/cache_utils.py b/src/transformers/cache_utils.py
index ac324ebb62..2b8b3f4f4e 100644
--- a/src/transformers/cache_utils.py
+++ b/src/transformers/cache_utils.py
@@ -1200,6 +1200,7 @@ class DynamicCache(Cache):
         offloading: bool = False,
         offload_only_non_sliding: bool = False,
     ):
+        self.is_stale = False
         layers = []
         # If a config is passed, use it to infer the layer types and initialize accordingly
         if config is not None:
diff --git a/src/transformers/models/cohere2/modular_cohere2.py b/src/transformers/models/cohere2/modular_cohere2.py
index d19055a1b7..1f02ad26ac 100644
--- a/src/transformers/models/cohere2/modular_cohere2.py
+++ b/src/transformers/models/cohere2/modular_cohere2.py
@@ -194,7 +194,15 @@ class Cohere2Attention(CohereAttention):
             query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

         if past_key_values is not None:
-            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
+            if isinstance(past_key_values, DynamicCache) and past_key_values.is_stale:
+                import warnings
+
+                warnings.warn(
+                    f"The cache is stale for layer {self.layer_idx}. Bypassing the internal cache update.",
+                    UserWarning,
+                )
+            else:
+                key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)

         attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
             self.config._attn_implementation, eager_attention_forward
diff --git a/src/transformers/models/deepseek_v3/modular_deepseek_v3.py b/src/transformers/models/deepseek_v3/modular_deepseek_v3.py
index 3c62a564a3..7ed3bd34a3 100644
--- a/src/transformers/models/deepseek_v3/modular_deepseek_v3.py
+++ b/src/transformers/models/deepseek_v3/modular_deepseek_v3.py
@@ -6,7 +6,7 @@ import torch.nn.functional as F
 from torch import nn

 from ... import initialization as init
-from ...cache_utils import Cache
+from ...cache_utils import Cache, DynamicCache
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import GenericForSequenceClassification, GenericForTokenClassification
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
@@ -256,7 +256,15 @@ class DeepseekV3Attention(nn.Module):
         key_states = torch.cat((k_pass, k_rot), dim=-1)

         if past_key_values is not None:
-            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
+            if isinstance(past_key_values, DynamicCache) and past_key_values.is_stale:
+                import warnings
+
+                warnings.warn(
+                    f"The cache is stale for layer {self.layer_idx}. Bypassing the internal cache update.",
+                    UserWarning,
+                )
+            else:
+                key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)

         if is_flash_attention_requested(self.config) and self.qk_head_dim != self.v_head_dim:
             value_states = F.pad(value_states, [0, self.qk_head_dim - self.v_head_dim])
diff --git a/src/transformers/models/gemma4/modular_gemma4.py b/src/transformers/models/gemma4/modular_gemma4.py
index a972738022..2b70f31332 100644
--- a/src/transformers/models/gemma4/modular_gemma4.py
+++ b/src/transformers/models/gemma4/modular_gemma4.py
@@ -985,12 +985,20 @@ class Gemma4TextAttention(nn.Module):
             value_states = value_states.transpose(1, 2)

         if past_key_values is not None:
-            if not self.is_kv_shared_layer:
-                key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
-            if self.store_full_length_kv:
-                if not hasattr(past_key_values, "shared_layers"):
-                    past_key_values.shared_layers = {}
-                past_key_values.shared_layers[self.layer_idx] = key_states, value_states
+            if isinstance(past_key_values, DynamicCache) and past_key_values.is_stale:
+                import warnings
+
+                warnings.warn(
+                    f"The cache is stale for layer {self.layer_idx}. Bypassing the internal cache update.",
+                    UserWarning,
+                )
+            else:
+                if not self.is_kv_shared_layer:
+                    key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
+                if self.store_full_length_kv:
+                    if not hasattr(past_key_values, "shared_layers"):
+                        past_key_values.shared_layers = {}
+                    past_key_values.shared_layers[self.layer_idx] = key_states, value_states

         attention_interface: Callable = eager_attention_forward
         if self.config._attn_implementation != "eager":
diff --git a/src/transformers/models/llama4/modeling_llama4.py b/src/transformers/models/llama4/modeling_llama4.py
index 08d50bd63f..ba642d5d83 100644
--- a/src/transformers/models/llama4/modeling_llama4.py
+++ b/src/transformers/models/llama4/modeling_llama4.py
@@ -389,7 +389,15 @@ class Llama4TextAttention(nn.Module):
         key_states = key_states.transpose(1, 2)

         if past_key_values is not None:
-            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
+            if isinstance(past_key_values, DynamicCache) and past_key_values.is_stale:
+                import warnings
+
+                warnings.warn(
+                    f"The cache is stale for layer {self.layer_idx}. Bypassing the internal cache update.",
+                    UserWarning,
+                )
+            else:
+                key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)

         attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
             self.config._attn_implementation, eager_attention_forward
diff --git a/src/transformers/models/ministral3/modular_ministral3.py b/src/transformers/models/ministral3/modular_ministral3.py
index 0a925f5d40..0a2bde879a 100644
--- a/src/transformers/models/ministral3/modular_ministral3.py
+++ b/src/transformers/models/ministral3/modular_ministral3.py
@@ -2,7 +2,7 @@ from collections.abc import Callable

 import torch

-from ...cache_utils import Cache
+from ...cache_utils import Cache, DynamicCache
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import (
     GenericForQuestionAnswering,
@@ -57,7 +57,15 @@ class Ministral3Attention(MistralAttention):
         ).to(query_states.dtype)

         if past_key_values is not None:
-            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
+            if isinstance(past_key_values, DynamicCache) and past_key_values.is_stale:
+                import warnings
+
+                warnings.warn(
+                    f"The cache is stale for layer {self.layer_idx}. Bypassing the internal cache update.",
+                    UserWarning,
+                )
+            else:
+                key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)

         attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
             self.config._attn_implementation, eager_attention_forward
diff --git a/src/transformers/models/mistral4/modular_mistral4.py b/src/transformers/models/mistral4/modular_mistral4.py
index acd9f1f601..5efbe66e83 100644
--- a/src/transformers/models/mistral4/modular_mistral4.py
+++ b/src/transformers/models/mistral4/modular_mistral4.py
@@ -18,7 +18,7 @@ import torch.nn.functional as F
 from torch import nn

 from ... import initialization as init
-from ...cache_utils import Cache
+from ...cache_utils import Cache, DynamicCache
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import GenericForSequenceClassification, GenericForTokenClassification
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
@@ -191,7 +191,15 @@ class Mistral4Attention(DeepseekV3Attention):
         ).to(query_states.dtype)

         if past_key_values is not None:
-            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
+            if isinstance(past_key_values, DynamicCache) and past_key_values.is_stale:
+                import warnings
+
+                warnings.warn(
+                    f"The cache is stale for layer {self.layer_idx}. Bypassing the internal cache update.",
+                    UserWarning,
+                )
+            else:
+                key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)

         if is_flash_attention_requested(self.config) and self.qk_head_dim != self.v_head_dim:
             value_states = F.pad(value_states, [0, self.qk_head_dim - self.v_head_dim])
diff --git a/src/transformers/models/olmo3/modular_olmo3.py b/src/transformers/models/olmo3/modular_olmo3.py
index 81325db7cc..11fb68b5ea 100644
--- a/src/transformers/models/olmo3/modular_olmo3.py
+++ b/src/transformers/models/olmo3/modular_olmo3.py
@@ -126,7 +126,15 @@ class Olmo3Attention(Olmo2Attention):
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

         if past_key_values is not None:
-            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
+            if isinstance(past_key_values, DynamicCache) and past_key_values.is_stale:
+                import warnings
+
+                warnings.warn(
+                    f"The cache is stale for layer {self.layer_idx}. Bypassing the internal cache update.",
+                    UserWarning,
+                )
+            else:
+                key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)

         attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
             self.config._attn_implementation, eager_attention_forward
diff --git a/src/transformers/models/qwen3/modular_qwen3.py b/src/transformers/models/qwen3/modular_qwen3.py
index 73cde6d89a..286d545e0c 100644
--- a/src/transformers/models/qwen3/modular_qwen3.py
+++ b/src/transformers/models/qwen3/modular_qwen3.py
@@ -17,7 +17,7 @@ from collections.abc import Callable

 import torch

-from ...cache_utils import Cache
+from ...cache_utils import Cache, DynamicCache
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_outputs import CausalLMOutputWithPast
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS
@@ -84,7 +84,15 @@ class Qwen3Attention(LlamaAttention):
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

         if past_key_values is not None:
-            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
+            if isinstance(past_key_values, DynamicCache) and past_key_values.is_stale:
+                import warnings
+
+                warnings.warn(
+                    f"The cache is stale for layer {self.layer_idx}. Bypassing the internal cache update.",
+                    UserWarning,
+                )
+            else:
+                key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)

         attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
             self.config._attn_implementation, eager_attention_forward