`_prepare_4d_attention_mask_for_sdpa` is not for causal attention but claims...

[minostauros]

... SDPA causal mask generation may be wrong for the mask generation.

transformers/src/transformers/modeling_attn_mask_utils.py

Lines 421 to 433 in 76fa17c

	if torch.all(mask == 1):
	if is_tracing:
	pass
	elif tgt_len == 1:
	# For query_length == 1, causal attention and bi-directional attention are the same.
	return None
	elif key_value_length == tgt_len:
	return None
	else:
	# Unfortunately, for query_length > 1 and key_value_length != query_length, we can not generally ignore the attention mask, as SDPA causal mask generation
	# may be wrong. We will set is_causal=False in SDPA and rely on Transformers attention_mask instead, hence not setting it to None here.
	# Reference: https://github.com/pytorch/pytorch/issues/108108
	return AttentionMaskConverter._expand_mask(mask=mask, dtype=dtype, tgt_len=tgt_len)

Will it be safe to just return None for the else: case?

For causal attention, we can just use _prepare_4d_causal_attention_mask_for_sdpa

Related issues:
pytorch/pytorch#108108
Dao-AILab/flash-attention@9e5e8bc
#28802

[amyeroberts]

cc @fxmarty

[fxmarty]

Hi @minostauros, thank you for the report.

... SDPA causal mask generation may be wrong for the mask generation.

_prepare_4d_attention_mask_for_sdpa does not handle causal masks. However,

Will it be safe to just return None for the else: case?

Yes, good catch, I'll fix that! This is a somewhat unlikely case though, where one would use past key values for typically encoder-type of models. How did you run into this case?

[minostauros]

This is a somewhat unlikely case though, where one would use past key values for typically encoder-type of models. How did you run into this case?

I didn't run into the specific section but I was just reviewing #28802 and was trying to add flash-attention-2 to BERT (BLIP-2 variant of BERT to be exact).
Thanks for confirmation!

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[minostauros]

This issue will be closed by #30138