torch.nn.TransformerEncoderLayer missing exception description information.

[triumph-wangyuyang]

🐛 Describe the bug

import torch
encoder_layer = torch.nn.TransformerEncoderLayer(d_model=52, nhead=1)
src = torch.rand(10, 32, 512)
out = encoder_layer(src)

When running this code, an AssertionError is thrown directly without giving effective information about the exception. The real error is that d_model is not equal to the third value in src. It is recommended to use the pytorch1.8.0 version. When this happens, throw Display the corresponding exception information.

Versions

pytorch: 1.8.0
Python version: 3.8
CUDA/cuDNN version: cuDNN 11.1
GPU models and configuration: RTX3060
Operating System：Windows

cc @albanD @mruberry @jbschlosser @walterddr @kshitij12345 @saketh-are @bhosmer @cpuhrsch @erichan1

[samdow]

Are you strongly attached to 1.8.0? On 1.13, it gives a more informative error:

  File "/Users/samdow/anaconda3/envs/1.13/lib/python3.10/site-packages/torch/nn/functional.py", line 5046, in multi_head_attention_forward
    assert embed_dim == embed_dim_to_check, \
AssertionError: was expecting embedding dimension of 52, but got 512

since d_model is the number of expected features of the input according to the docs, it and the final dim of src must match

[triumph-wangyuyang]

TransformerEncoderLayer

Currently doing a test experiment, need to use pytorch1.8 version, and then use TransformerEncoderLayer, this problem is triggered, I feel very confused, I will test with a higher version later, and found that the higher version has provided the corresponding information description, thank you .