Update handling for KeyError in state_dict.pop() for non-existing keys.
Changed state_dict.pop(f"h.{d}.attn.bias") to state_dict.pop(f"h.{d}.attn.bias", None) to prevent KeyError exceptions.
The following code can re-produce the issue
```
from transformers import AutoTokenizer, GPT2Model, GPT2Config
from flash_attn.models.gpt import GPTLMHeadModel, GPTModel
# >>> transformers.__version__
# '4.38.2'
model_path = 'gpt2'
output_model_path = 'gpt2_model'
config = GPT2Config.from_pretrained(model_path, output_hidden_states=True)
model = GPT2Model.from_pretrained(model_path, from_tf=False, config=config)
'''
model fine-tuning here
'''
# dump the fine-tuned model
model.save_pretrained(output_model_path)
# load the fine-tuned model
config = GPT2Config.from_pretrained(output_model_path, output_hidden_states=True)
model = GPTModel.from_pretrained(output_model_path, config=config, strict=True) # failed due to KeyError: 'h.0.attn.bias'
model = GPTLMHeadModel.from_pretrained(output_model_path, config=config, strict=True) # failed due to KeyError: 'h.0.attn.bias'
```
0d810cfb73