'attention.bias' and 'attention.masked_bias' not in hf_layer.state_dict() when converting gpt-neox model to huggingface

[johntzwei]

Describe the bug
A clear and concise description of what the bug is.

I encounter the following error when I am converting GPTNeoX models to Huggingface using the tools/convert_module_to_hf.py script.

(gpt-neox) johnny@ink-lucy:~/gpt-neox$ bash haveibeentrainedon/wikitext/pilot/convert_to_hf.sh 
[2023-08-18 23:37:21,695] [INFO] [real_accelerator.py:133:get_accelerator] Setting ds_accelerator to cuda (auto detect)
> building GPT2BPETokenizer tokenizer ...
 > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304)
Saving weights in fp16 precision...
  0%|                                                                                           | 0/24 [00:00<?, ?it/s]
Traceback (most recent call last):
  File "./tools/convert_module_to_hf.py", line 307, in <module>
    hf_model = convert(args.input_dir, loaded_config, args.output_dir)
  File "./tools/convert_module_to_hf.py", line 230, in convert
    state_dict["attention.bias"] = hf_layer.state_dict()["attention.bias"]
KeyError: 'attention.bias'

Expected behavior
Successful conversion.

Proposed solution
If you comment out lines 230 and 231, the script will run through. From an eyeballing of the results, it doesn't seem like language modelling performance seriously degraded. Could this be some code that was supposed to be taken out?

Additional context
This is for a model trained with the config configs/pythia/410m.yml

[shuheikurita]

Use pip install transformers==4.30.2

transformers>=4.31.0:
ipdb> hf_layer.state_dict().keys()
odict_keys(['input_layernorm.weight', 'input_layernorm.bias', 'post_attention_layernorm.weight', 'post_attention_layernorm.bias', 'attention.rotary_emb.inv_freq', 'attention.query_key_value.weight', 'attention.query_key_value.bias', 'attention.dense.weight', 'attention.dense.bias', 'mlp.dense_h_to_4h.weight', 'mlp.dense_h_to_4h.bias', 'mlp.dense_4h_to_h.weight', 'mlp.dense_4h_to_h.bias'])

transformers<=4.30.2 :
ipdb> hf_layer.state_dict().keys()
odict_keys(['input_layernorm.weight', 'input_layernorm.bias', 'post_attention_layernorm.weight', 'post_attention_layernorm.bias', 'attention.bias', 'attention.masked_bias', 'attention.rotary_emb.inv_freq', 'attention.query_key_value.weight', 'attention.query_key_value.bias', 'attention.dense.weight', 'attention.dense.bias', 'mlp.dense_h_to_4h.weight', 'mlp.dense_h_to_4h.bias', 'mlp.dense_4h_to_h.weight', 'mlp.dense_4h_to_h.bias'])

You can find missing 'attention.bias', 'attention.masked_bias' before transformers==4.30.2.

[dashstander]

Thanks so much @shuheikurita, I just made a PR to update the transformers version