GitLab

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'

```

Co-authored-by: danthe3rd <danthe3rd>

* q

* add comment.

* fix name.

* set inv function.

* add map back function.

* handle gqa.

* add type annotation to avoid confusion.

* fix docstr.

* test inverse remap logic.

* uneql rank.

* trim.

* enable passing in number of heads for each rank.

* simplify.

* simplify.

* cleanup.

* fix col parallel.

* fix bug with row parallel.

* fit out proj.

* refac.

* fix sharding logic.

* refac sharding.

* refac.

* support multiple of.

* make fn reuseable.

* fix bug in dimensions.

* scaffold.

* test uneven heads.

* fix test by adding barrier.

* refac.

* reuse code.

* clean up.