Using max_position_embeddings instead of max_sequence_length to standardise with HF

Fix causal mode

fixed NaNs

Update modeling_gptbert.py

make FlashAttention logic more robust

fix

removed SDPA

Update modeling_gptbert.py

fixed SDPA for older PyTorch versions

fix

Update modeling_gptbert.py

Update modeling_gptbert.py

fix

fix

fix

fix attention mask

fix

fix

Update modeling_gptbert.py

Update modeling_gptbert.py

fixes and optimizations

Fix import error

Update modeling_gptbert.py

Update modeling_gptbert.py

Update modeling_gptbert.py

Update modeling_gptbert.py

Update modeling_gptbert.py

Update modeling_gptbert.py

Update modeling_gptbert.py

Update modeling_gptbert.py

Checking for errors in the CastedLinear

Update modeling_gptbert.py

More testing

Add checking for NaN

Removing added print

Testing output

fix flasshattention

test

test

fix attention masks

fix import error

Fix error

Add FlashAttention + unpadding support

Update modeling_gptbert.py

updated MLM predictions

Pass throught _init_weights

Fixed classifier head

Upload folder using huggingface_hub