CANINE unexpectedly requires input_ids anyway

[marcmk6]

System Info

• transformers version: 4.33.2
• Platform: Linux-5.15.109+-x86_64-with-glibc2.35
• Python version: 3.10.12
• Huggingface_hub version: 0.17.2
• Safetensors version: 0.3.3
• Accelerate version: not installed
• Accelerate config: not found
• PyTorch version (GPU?): 2.0.1+cu118 (False)
• Tensorflow version (GPU?): 2.13.0 (False)
• Flax version (CPU?/GPU?/TPU?): 0.7.2 (cpu)
• Jax version: 0.4.14
• JaxLib version: 0.4.14
• Using GPU in script?:
• Using distributed or parallel set-up in script?:

Who can help?

@ArthurZucker and @younesbelkada

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

from transformers import CanineModel, BertModel
import torch

BERT_model = BertModel.from_pretrained('bert-base-uncased')
canine_model = CanineModel.from_pretrained('google/canine-c')

fake_input = torch.rand(1, 10, 768)

_ = BERT_model.forward(inputs_embeds=fake_input) # no error
_ = canine_model.forward(inputs_embeds=fake_input) # error

The error message

File /miniconda3/envs/tmp/lib/python3.10/site-packages/transformers/models/canine/modeling_canine.py:1172, in CanineModel.forward(self, input_ids, attention_mask, token_type_ids, position_ids, head_mask, inputs_embeds, output_attentions, output_hidden_states, return_dict)
   1162 input_char_embeddings = self.char_embeddings(
   1163     input_ids=input_ids,
   1164     position_ids=position_ids,
   1165     token_type_ids=token_type_ids,
   1166     inputs_embeds=inputs_embeds,
   1167 )
   1169 # Contextualize character embeddings using shallow Transformer.
   1170 # We use a 3D attention mask for the local attention.
   1171 # `input_char_encoding`: shape (batch_size, char_seq_len, char_dim)
-> 1172 char_attention_mask = self._create_3d_attention_mask_from_input_mask(input_ids, attention_mask)
   1173 init_chars_encoder_outputs = self.initial_char_encoder(
   1174     input_char_embeddings,
   1175     attention_mask=char_attention_mask,
   1176     output_attentions=output_attentions,
   1177     output_hidden_states=output_hidden_states,
   1178 )
   1179 input_char_encoding = init_chars_encoder_outputs.last_hidden_state

File /miniconda3/envs/tmp/lib/python3.10/site-packages/transformers/models/canine/modeling_canine.py:1042, in CanineModel._create_3d_attention_mask_from_input_mask(self, from_tensor, to_mask)
   1031 def _create_3d_attention_mask_from_input_mask(self, from_tensor, to_mask):
   1032     """
   1033     Create 3D attention mask from a 2D tensor mask.
   1034 
   (...)
   1040         float Tensor of shape [batch_size, from_seq_length, to_seq_length].
   1041     """
-> 1042     batch_size, from_seq_length = from_tensor.shape[0], from_tensor.shape[1]
   1044     to_seq_length = to_mask.shape[1]
   1046     to_mask = torch.reshape(to_mask, (batch_size, 1, to_seq_length)).float()

AttributeError: 'NoneType' object has no attribute 'shape'

Expected behavior

According to doc, the forward should work with either input_ids or inputs_embeds provided. But it turns out input_ids is used for deriving other variables in the code in all cases.