Fix: Import structure cleaning with Ruff.

src/transformers/__init__.py

@@ -570,7 +570,6 @@
     "models.m2m_100": ["M2M100Config"],
     "models.mamba": ["MambaConfig"],
     "models.mamba2": ["Mamba2Config"],
-    "models.xlstm": ["xLSTMConfig"],
     "models.marian": ["MarianConfig"],
     "models.markuplm": [
         "MarkupLMConfig",
@@ -869,6 +868,7 @@
     "models.xlm_roberta": ["XLMRobertaConfig"],
     "models.xlm_roberta_xl": ["XLMRobertaXLConfig"],
     "models.xlnet": ["XLNetConfig"],
+    "models.xlstm": ["xLSTMConfig"],
     "models.xmod": ["XmodConfig"],
     "models.yolos": ["YolosConfig"],
     "models.yoso": ["YosoConfig"],
@@ -2732,13 +2732,6 @@
             "Mamba2PreTrainedModel",
         ]
     )
-    _import_structure["models.xlstm"].extend(
-        [
-            "xLSTMForCausalLM",
-            "xLSTMModel",
-            "xLSTMPreTrainedModel",
-        ]
-    )
     _import_structure["models.marian"].extend(
         ["MarianForCausalLM", "MarianModel", "MarianMTModel", "MarianPreTrainedModel"]
     )
@@ -3861,6 +3854,13 @@
             "load_tf_weights_in_xlnet",
         ]
     )
+    _import_structure["models.xlstm"].extend(
+        [
+            "xLSTMForCausalLM",
+            "xLSTMModel",
+            "xLSTMPreTrainedModel",
+        ]
+    )
     _import_structure["models.xmod"].extend(
         [
             "XmodForCausalLM",

src/transformers/models/xlstm/configuration_xlstm.py

@@ -14,10 +14,9 @@
 # limitations under the License.
 """XLSTM configuration"""
 
-
-
 from ...configuration_utils import PretrainedConfig
-from ...utils import logging, is_xlstm_available
+from ...utils import is_xlstm_available, logging
+
 
 if is_xlstm_available():
     from xlstm.xlstm_large.model import (

src/transformers/models/xlstm/modeling_xlstm.py

@@ -8,8 +8,6 @@
 from torch import nn
 from torch.nn import CrossEntropyLoss
 
-
-
 from ...generation import GenerationMixin
 from ...modeling_utils import PreTrainedModel
 from ...utils import (
@@ -21,13 +19,14 @@
     is_xlstm_available,
 )
 
+
 if is_xlstm_available():
     from xlstm.xlstm_large.model import (
         RMSNorm,
         mLSTMBlock,
         mLSTMStateType,
         soft_cap,
-    )    
+    )
 else:
     mLSTMBlock = None
 
@@ -421,7 +420,11 @@ def forward(
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
         inserted_bos_token = False
-        if (cache_params is None or cache_params.rnn_state_initial) and self.config.force_bos_token_insert and input_ids is not None:
+        if (
+            (cache_params is None or cache_params.rnn_state_initial)
+            and self.config.force_bos_token_insert
+            and input_ids is not None
+        ):
             if not is_torchdynamo_compiling():
                 if bool(torch.all(input_ids[0, 0] != self.config.bos_token_id).cpu()):
                     input_ids = torch.cat(
@@ -444,9 +447,21 @@ def forward(
         if inserted_bos_token:
             hidden_states = hidden_states[:, 1:]
             if hasattr(xlstm_outputs, "hidden_states"):
-                xlstm_outputs_mod = (hidden_states, xlstm_outputs.cache_params, tuple(hidden_state[:, 1:] for hidden_state in xlstm_outputs.hidden_states) if xlstm_outputs.hidden_states is not None else None)
-            elif len(xlstm_outputs) == 3: 
-                xlstm_outputs_mod = (hidden_states, xlstm_outputs[1], tuple(hidden_state[:, 1:] for hidden_state in xlstm_outputs[2]) if xlstm_outputs[2] is not None else None)
+                xlstm_outputs_mod = (
+                    hidden_states,
+                    xlstm_outputs.cache_params,
+                    tuple(hidden_state[:, 1:] for hidden_state in xlstm_outputs.hidden_states)
+                    if xlstm_outputs.hidden_states is not None
+                    else None,
+                )
+            elif len(xlstm_outputs) == 3:
+                xlstm_outputs_mod = (
+                    hidden_states,
+                    xlstm_outputs[1],
+                    tuple(hidden_state[:, 1:] for hidden_state in xlstm_outputs[2])
+                    if xlstm_outputs[2] is not None
+                    else None,
+                )
             elif len(xlstm_outputs) == 2:
                 xlstm_outputs_mod = (hidden_states, xlstm_outputs[1])
             elif len(xlstm_outputs) == 1:

tests/models/xlstm/test_modeling_xlstm.py

@@ -35,7 +35,7 @@
         xLSTMForCausalLM,
         xLSTMModel,
     )
-    from transformers.models.xlstm.modeling_xlstm import xLSTMCache, mLSTMBlock
+    from transformers.models.xlstm.modeling_xlstm import mLSTMBlock, xLSTMCache
     from transformers.pytorch_utils import is_torch_greater_or_equal_than_2_2
 else:
     is_torch_greater_or_equal_than_2_2 = False
@@ -53,7 +53,7 @@ def __init__(
         vocab_size=99,
         embedding_dim=128,
         qk_dim_factor=0.5,
-        v_dim_factor=1.,
+        v_dim_factor=1.0,
         num_blocks=2,
         max_position_embeddings=512,
         type_vocab_size=16,
@@ -94,17 +94,14 @@ def __init__(
         self.num_hidden_layers = self.num_blocks
         self.hidden_size = self.embedding_dim
 
-
     def get_large_model_config(self):
         cfg = xLSTMConfig.from_pretrained("NX-AI/xLSTM-7b")
         # this is needed for compatibility with generic tests
         cfg.hidden_size = cfg.embedding_dim
         cfg.num_hidden_layers = cfg.num_blocks
         return cfg
 
-    def prepare_config_and_inputs(
-        self, scale_attn_by_inverse_layer_idx=False, reorder_and_upcast_attn=False
-    ):
+    def prepare_config_and_inputs(self, scale_attn_by_inverse_layer_idx=False, reorder_and_upcast_attn=False):
         input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
 
         sequence_labels = None