causal_mask in GPT2Attention should not be broadcastable across the seq_len

### System Info Python : 3.8.2 torch : 2.2.0.dev20231207+cu121 transformers : 4.31.0 torchvision : 0.17.0.dev20231207+cu121 cuda version : 12.1 In `transformers.models.gpt2.modeling_gpt2.GPT2Attention` https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L202 the `causal_mask` must have the same shape for the last 2 dims, otherwise if the `max_position_embeddings=1` while the sequence length is longer than 1, the resulted attention weights leads to attending the future tokens. See the steps to reproduce the behavior for details. Normally, one wouldn't set `max_position_embeddings=1`, but nevertheless the broadcasting should not happen. ### Who can help? _No response_ ### 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 Code to reproduce the issue: ``` import torch import transformers import matplotlib.pyplot as plt from mpl_toolkits.axes_grid1 import make_axes_locatable bsz, seq_len, hid = 2, 3, 4 fig, axes = plt.subplots(ncols=3, figsize=(9,2)) for n_positions, ax in zip([1, 2, seq_len], axes): attn = transformers.models.gpt2.modeling_gpt2.GPT2Attention(transformers.GPT2Config(n_embd=hid, n_layer=1, n_head=1, n_positions=n_positions)) ax.axis(False) ax.set_title('attn_weights, n_positions=%d' % n_positions, fontsize=9) attn_input = torch.randn(bsz, seq_len, hid) try: attn_output, _, attn_weights = attn(attn_input, output_attentions=True) except Exception as e: print('n_positions=%d' % n_positi…