File transformers/models/modernbert/modeling_modernbert.py:472, in ModernBertModel.forward(self, input_ids, attention_mask, position_ids, inputs_embeds, **kwargs)
    465 if not isinstance(attention_mask_mapping := attention_mask, dict):
    466     mask_kwargs = {
    467         "config": self.config,
    468         "inputs_embeds": hidden_states,
    469         "attention_mask": attention_mask,
    470     }
    471     attention_mask_mapping = {
--> 472         "full_attention": create_bidirectional_mask(**mask_kwargs),
    473         "sliding_attention": create_bidirectional_sliding_window_mask(**mask_kwargs),
    474     }
    476 position_embeddings = {}
    477 for layer_type in set(self.config.layer_types):

File transformers/utils/deprecation.py:171, in deprecate_kwarg.<locals>.wrapper.<locals>.wrapped_func(*args, **kwargs)
    167 elif minimum_action in (Action.NOTIFY, Action.NOTIFY_ALWAYS) and not is_torchdynamo_compiling():
    168     # DeprecationWarning is ignored by default, so we use FutureWarning instead
    169     warnings.warn(message, FutureWarning, stacklevel=2)
--> 171 return func(*args, **kwargs)

File transformers/masking_utils.py:1071, in create_bidirectional_mask(config, inputs_embeds, attention_mask, encoder_hidden_states, past_key_values, or_mask_function, and_mask_function)
   1068     use_vmap = True
   1070 # We now create the mask
-> 1071 attention_mask = mask_interface(
   1072     batch_size=batch_size,
   1073     q_length=q_length,
   1074     kv_length=kv_length,
   1075     q_offset=q_offset,
   1076     kv_offset=kv_offset,
   1077     mask_function=mask_factory_function,
   1078     attention_mask=attention_mask,
   1079     # Additional kwargs for sdpa
   1080     allow_is_causal_skip=False,
   1081     allow_is_bidirectional_skip=allow_is_bidirectional_skip,
   1082     dtype=dtype,  # Additional kwarg for eager
   1083     config=config,  # Pass the config as well, in case someone wants to easily have their own mask_interface
   1084     use_vmap=use_vmap,  # Short-circuit to non-vmap expansions for the mask
   1085     device=device,
   1086 )
   1087 return attention_mask

File transformers/masking_utils.py:492, in sdpa_mask(batch_size, q_length, kv_length, q_offset, kv_offset, mask_function, attention_mask, local_size, allow_is_causal_skip, allow_is_bidirectional_skip, allow_torch_fix, use_vmap, device, **kwargs)
    487 if isinstance(q_length, torch.Tensor):
    488     logger.warning_once(
    489         "`cache_position` is deprecated as an arg, and will be removed in Transformers v5.6. Please use `q_length` and "
    490         "`q_offset` instead, similarly to `kv_length` and `kv_offset`"
    491     )
--> 492     q_length, q_offset = q_length.shape[0], q_length[0].to(device)
    494 # Potentially pad the 2D mask
    495 padding_mask = prepare_padding_mask(attention_mask, kv_length, kv_offset)

IndexError: tuple index out of range

---

import torch
from transformers import AutoModel, AutoTokenizer

# Note that using bert-base-uncased creates another separate ERROR
path = "answerdotai/ModernBERT-base"

model = AutoModel.from_pretrained(path)
tokenizer = AutoTokenizer.from_pretrained(path)

# Generate dummy inputs
dummy = dict(tokenizer(["test inputs"], return_tensors="pt"))

torch.onnx.export(
    model,
    (dummy,),
    dynamo=False
)

---

diff --git a/transformers/masking_utils.py b/tmp/patch.py
index 45e43fd..3d9f496 100644
--- a/transformers/masking_utils.py
+++ b/tmp/patch.py
@@ -484,7 +484,7 @@ def sdpa_mask(
 
     """
     # For BC on `cache_positions` that used to be an arg at the position of `q_length`
-    if isinstance(q_length, torch.Tensor):
+    if isinstance(q_length, torch.Tensor) and q_length.ndim > 0:
         logger.warning_once(
             "`cache_position` is deprecated as an arg, and will be removed in Transformers v5.6. Please use `q_length` and "
             "`q_offset` instead, similarly to `kv_length` and `kv_offset`"
@@ -689,7 +689,7 @@ def flex_attention_mask(
             An optional device to create the mask on.
     """
     # For BC on `cache_positions` that used to be an arg at the position of `q_length`
-    if isinstance(q_length, torch.Tensor):
+    if isinstance(q_length, torch.Tensor) and q_length.ndim > 0:
         logger.warning_once(
             "`cache_position` is deprecated as an arg, and will be removed in Transformers v5.6. Please use `q_length` and "
             "`q_offset` instead, similarly to `kv_length` and `kv_offset`"