Support passing flash_attn_kwargs when gradient_checkpointing is enabled (#37037)

efsotr · web-flow · commit 2b4734bd4907 · 2025-03-31T10:53:02.000+02:00
* support passing flash_attn_kwargs when gradient_checkpointing is enabled

* make modeling_deepspeek_v3.py consistent with modular_deepseek_v3.py
diff --git a/examples/modular-transformers/modeling_dummy.py b/examples/modular-transformers/modeling_dummy.py
@@ -4,6 +4,7 @@
 #             the file from the modular. If any change should be done, please apply the change to the
 #                          modular_dummy.py file directly. One of our CI enforces this.
 #                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+from functools import partial
 from typing import Callable, Optional, Tuple, Union
 
 import torch
@@ -544,7 +545,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     causal_mask,
                     position_ids,
diff --git a/examples/modular-transformers/modeling_multimodal1.py b/examples/modular-transformers/modeling_multimodal1.py
@@ -4,6 +4,7 @@
 #             the file from the modular. If any change should be done, please apply the change to the
 #                          modular_multimodal1.py file directly. One of our CI enforces this.
 #                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+from functools import partial
 from typing import Callable, Optional, Tuple, Union
 
 import torch
@@ -544,7 +545,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     causal_mask,
                     position_ids,
diff --git a/src/transformers/models/aria/modeling_aria.py b/src/transformers/models/aria/modeling_aria.py
@@ -19,6 +19,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from dataclasses import dataclass
+from functools import partial
 from typing import Callable, List, Optional, Tuple, Union
 
 from ...activations import ACT2FN
@@ -963,7 +964,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     causal_mask,
                     position_ids,
diff --git a/src/transformers/models/cohere/modeling_cohere.py b/src/transformers/models/cohere/modeling_cohere.py
@@ -27,6 +27,7 @@
 # This file is based on the LLama model definition file in transformers
 
 
+from functools import partial
 from typing import Callable, List, Optional, Tuple, Union
 
 import torch
@@ -613,7 +614,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     causal_mask,
                     position_ids,
diff --git a/src/transformers/models/cohere2/modeling_cohere2.py b/src/transformers/models/cohere2/modeling_cohere2.py
@@ -19,6 +19,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+from functools import partial
 from typing import Callable, List, Optional, Tuple, Union
 
 import torch
@@ -634,7 +635,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     position_embeddings,
                     causal_mask,
diff --git a/src/transformers/models/cohere2/modular_cohere2.py b/src/transformers/models/cohere2/modular_cohere2.py
@@ -13,6 +13,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+from functools import partial
 from typing import Callable, Optional, Tuple, Union
 
 import torch
@@ -533,7 +534,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     position_embeddings,
                     causal_mask,
diff --git a/src/transformers/models/deepseek_v3/modeling_deepseek_v3.py b/src/transformers/models/deepseek_v3/modeling_deepseek_v3.py
@@ -5,6 +5,7 @@
 #                          modular_deepseek_v3.py file directly. One of our CI enforces this.
 #                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
 import math
+from functools import partial
 from typing import Callable, Optional, Tuple, Union
 
 import torch
@@ -759,7 +760,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     causal_mask,
                     position_ids,
diff --git a/src/transformers/models/diffllama/modeling_diffllama.py b/src/transformers/models/diffllama/modeling_diffllama.py
@@ -22,6 +22,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import math
+from functools import partial
 from typing import Optional, Tuple, Union
 
 import torch
@@ -852,7 +853,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     causal_mask,
                     position_ids,
diff --git a/src/transformers/models/emu3/modeling_emu3.py b/src/transformers/models/emu3/modeling_emu3.py
@@ -21,7 +21,7 @@
 # limitations under the License.
 
 import math
-from functools import cached_property
+from functools import cached_property, partial
 from typing import Callable, List, Optional, Tuple, Union
 
 import torch
@@ -1439,7 +1439,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     causal_mask,
                     position_ids,
diff --git a/src/transformers/models/gemma2/modeling_gemma2.py b/src/transformers/models/gemma2/modeling_gemma2.py
@@ -19,6 +19,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+from functools import partial
 from typing import Callable, Optional, Tuple, Union
 
 import torch
@@ -645,7 +646,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     position_embeddings,
                     causal_mask,
diff --git a/src/transformers/models/gemma2/modular_gemma2.py b/src/transformers/models/gemma2/modular_gemma2.py
@@ -13,6 +13,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+from functools import partial
 from typing import Callable, Optional, Tuple, Union
 
 import torch
@@ -491,7 +492,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     position_embeddings,
                     causal_mask,
diff --git a/src/transformers/models/gemma3/modeling_gemma3.py b/src/transformers/models/gemma3/modeling_gemma3.py
@@ -22,6 +22,7 @@
 import copy
 from collections.abc import Callable
 from dataclasses import dataclass
+from functools import partial
 from typing import List, Optional, Tuple, Union
 
 import torch
@@ -732,7 +733,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     position_embeddings_global,
                     position_embeddings_local,
diff --git a/src/transformers/models/gemma3/modular_gemma3.py b/src/transformers/models/gemma3/modular_gemma3.py
@@ -16,6 +16,7 @@
 import copy
 from collections.abc import Callable
 from dataclasses import dataclass
+from functools import partial
 from typing import Any, Dict, List, Optional, Tuple, Union
 
 import torch
@@ -662,7 +663,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     position_embeddings_global,
                     position_embeddings_local,
diff --git a/src/transformers/models/glm/modeling_glm.py b/src/transformers/models/glm/modeling_glm.py
@@ -19,6 +19,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+from functools import partial
 from typing import Callable, Optional, Tuple, Union
 
 import torch
@@ -594,7 +595,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     causal_mask,
                     position_ids,
diff --git a/src/transformers/models/granite/modeling_granite.py b/src/transformers/models/granite/modeling_granite.py
@@ -19,6 +19,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+from functools import partial
 from typing import Callable, List, Optional, Tuple, Union
 
 import torch
@@ -593,7 +594,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     causal_mask,
                     position_ids,
diff --git a/src/transformers/models/granite/modular_granite.py b/src/transformers/models/granite/modular_granite.py
@@ -13,6 +13,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+from functools import partial
 from typing import List, Optional, Tuple, Union
 
 import torch
@@ -185,7 +186,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     causal_mask,
                     position_ids,
diff --git a/src/transformers/models/helium/modeling_helium.py b/src/transformers/models/helium/modeling_helium.py
@@ -20,6 +20,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import math
+from functools import partial
 from typing import Callable, Optional, Tuple, Union
 
 import torch
@@ -581,7 +582,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     causal_mask,
                     position_ids,
diff --git a/src/transformers/models/llama/modeling_llama.py b/src/transformers/models/llama/modeling_llama.py
@@ -17,6 +17,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+from functools import partial
 from typing import Callable, Optional, Tuple, Union
 
 import torch
@@ -583,7 +584,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     causal_mask,
                     position_ids,
diff --git a/src/transformers/models/mistral/modeling_mistral.py b/src/transformers/models/mistral/modeling_mistral.py
@@ -4,6 +4,7 @@
 #             the file from the modular. If any change should be done, please apply the change to the
 #                          modular_mistral.py file directly. One of our CI enforces this.
 #                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+from functools import partial
 from typing import Callable, List, Optional, Tuple, Union
 
 import torch
@@ -548,7 +549,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     causal_mask,
                     position_ids,
diff --git a/src/transformers/models/mixtral/modeling_mixtral.py b/src/transformers/models/mixtral/modeling_mixtral.py
@@ -24,6 +24,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from functools import partial
 from typing import Callable, List, Optional, Tuple, Union
 
 import torch
@@ -672,7 +673,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     causal_mask,
                     position_ids,
diff --git a/src/transformers/models/mixtral/modular_mixtral.py b/src/transformers/models/mixtral/modular_mixtral.py
@@ -19,6 +19,7 @@
 # limitations under the License.
 """PyTorch Mixtral model."""
 
+from functools import partial
 from typing import List, Optional, Tuple, Union
 
 import torch
@@ -400,7 +401,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     causal_mask,
                     position_ids,
diff --git a/src/transformers/models/moonshine/modeling_moonshine.py b/src/transformers/models/moonshine/modeling_moonshine.py
@@ -18,6 +18,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from functools import partial
 from typing import Callable, Optional, Tuple, Union
 
 import numpy as np
@@ -936,7 +937,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     causal_mask,
                     encoder_hidden_states,
diff --git a/src/transformers/models/moonshine/modular_moonshine.py b/src/transformers/models/moonshine/modular_moonshine.py
@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from functools import partial
 from typing import Callable, Optional, Tuple, Union
 
 import torch
@@ -832,7 +833,7 @@ def forward(
 
             if self.gradient_checkpointing and self.training:
                 layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
                     hidden_states,
                     causal_mask,
                     encoder_hidden_states,
diff --git a/src/transformers/models/olmo/modeling_olmo.py b/src/transformers/models/olmo/modeling_olmo.py
diff --git a/src/transformers/models/olmo2/modeling_olmo2.py b/src/transformers/models/olmo2/modeling_olmo2.py
diff --git a/src/transformers/models/phi/modeling_phi.py b/src/transformers/models/phi/modeling_phi.py
diff --git a/src/transformers/models/phi/modular_phi.py b/src/transformers/models/phi/modular_phi.py
diff --git a/src/transformers/models/phi3/modeling_phi3.py b/src/transformers/models/phi3/modeling_phi3.py
diff --git a/src/transformers/models/qwen2/modeling_qwen2.py b/src/transformers/models/qwen2/modeling_qwen2.py
