Fix transformers 5.6 remote-code load

- modeling_kimi_k25.py: add direct relative import of configuration_deepseek so transformers 5.6 fetches the 2-level-deep file into the dynamic-module cache (get_class_from_dynamic_module previously failed with FileNotFoundError).
- modeling_deepseek.py: drop is_torch_fx_available import + dead fx-wrap block (symbol removed in transformers 5.x).

modeling_deepseek.py

@@ -44,7 +44,6 @@ from transformers.utils import (add_start_docstrings,
                                 is_flash_attn_2_available,
                                 is_flash_attn_greater_or_equal_2_10, logging,
                                 replace_return_docstrings)
-from transformers.utils.import_utils import is_torch_fx_available
 
 from .configuration_deepseek import DeepseekV3Config
 
@@ -55,12 +54,6 @@ if is_flash_attn_2_available():
 
 # This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
 # It means that the function will not be traced through and simply appear as a node in the graph.
-if is_torch_fx_available():
-    if not is_torch_greater_or_equal_than_1_13:
-        import torch.fx
-
-    _prepare_4d_causal_attention_mask = torch.fx.wrap(
-        _prepare_4d_causal_attention_mask)
 
 logger = logging.get_logger(__name__)
 

modeling_kimi_k25.py

@@ -64,6 +64,7 @@ from transformers.models.llava.modeling_llava import \
 from transformers.utils import is_flash_attn_2_available
 
 from .configuration_kimi_k25 import KimiK25Config
+from .configuration_deepseek import DeepseekV3Config  # noqa: F401  (ensure dynamic-module fetch)
 from .modeling_deepseek import DeepseekV3ForCausalLM
 
 # Flash attention imports