update modular files to match source

Author: kallewoof

src/transformers/models/olmo_hybrid/modeling_olmo_hybrid.py

@@ -38,7 +38,7 @@
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
+from ...utils import TransformersKwargs, agnostic, auto_docstring, can_return_tuple, logging
 from ...utils.generic import maybe_autocast, merge_with_config_defaults
 from ...utils.import_utils import is_flash_linear_attention_available
 from ...utils.output_capturing import capture_outputs
@@ -695,7 +695,7 @@ def __init__(self, config: OlmoHybridConfig, layer_idx: int):
             else FusedRMSNormGated(
                 self.head_v_dim,
                 eps=1e-5,
-                device=torch.cuda.current_device(),
+                device=agnostic.gpu.current_device(),
                 dtype=config.dtype if config.dtype is not None else torch.get_default_dtype(),
             )
         )

src/transformers/models/qwen3_5_moe/modeling_qwen3_5_moe.py

@@ -45,7 +45,7 @@
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging, torch_compilable_check
+from ...utils import TransformersKwargs, agnostic, auto_docstring, can_return_tuple, logging, torch_compilable_check
 from ...utils.generic import is_flash_attention_requested, maybe_autocast, merge_with_config_defaults
 from ...utils.import_utils import is_causal_conv1d_available, is_flash_linear_attention_available
 from ...utils.output_capturing import OutputRecorder, capture_outputs
@@ -396,7 +396,7 @@ def __init__(self, config: Qwen3_5MoeConfig, layer_idx: int):
                 self.head_v_dim,
                 eps=self.layer_norm_epsilon,
                 activation=self.activation,
-                device=torch.cuda.current_device(),
+                device=agnostic.gpu.current_device(),
                 dtype=config.dtype if config.dtype is not None else torch.get_default_dtype(),
             )
         )

src/transformers/models/qwen3_next/modeling_qwen3_next.py

@@ -42,7 +42,7 @@
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
+from ...utils import TransformersKwargs, agnostic, auto_docstring, can_return_tuple, logging
 from ...utils.generic import maybe_autocast, merge_with_config_defaults
 from ...utils.import_utils import is_causal_conv1d_available, is_flash_linear_attention_available
 from ...utils.output_capturing import OutputRecorder, capture_outputs
@@ -541,7 +541,7 @@ def __init__(self, config: Qwen3NextConfig, layer_idx: int):
                 self.head_v_dim,
                 eps=self.layer_norm_epsilon,
                 activation=self.activation,
-                device=torch.cuda.current_device(),
+                device=agnostic.gpu.current_device(),
                 dtype=config.dtype if config.dtype is not None else torch.get_default_dtype(),
             )
         )