[YHC66]

Description

AttentionBridge._apply_reconstruct_attention_mask treats every 4D mask as already encoding causal semantics and skips the local causal mask. That holds for the [batch, 1, seq, seq] masks most models build, but HF's OpenAIGPTModel passes a padding-only [batch, 1, 1, seq] mask and leaves the causal part to the attention module's bias buffer. So on GPT-1, any attention mask lets every position attend to later tokens.

This is reachable by passing attention_mask= or padding_side="left" to the bridge. Without a mask, and for right-padded string lists (which pass no mask), GPT-1 was already correct.

On openai-community/openai-gpt with "the capital of france is" and an all-ones mask:

	Before	After
Max logit difference from HF	20.59	2.1e-5 (same as with no mask)
Top-1 next token	all (HF: a)	a
Attention from position 0 to later tokens, block 0	~0.78	0

The fix only treats a 4D mask as authoritative when it has its own query rows (shape[-2] != 1). A single query row is also what a causal 4D mask looks like on a cached decode step, but the causal mask allows every key for that row, so applying it there changes nothing; there is a test for that case.

I found this with a padding comparison across the supported causal LM architectures, following #1952. GPT-1 was the only one showing it, and it is the only supported architecture I found in transformers that passes this mask shape.

Not related to #1969, which is about hook_out on the same model.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Testing

• New tests in test_mask_slicing.py and test_openai_gpt_adapter.py (the GPT-1 ones use a tiny random model, no download): 4 fail before the fix, all 12 pass after.
• tests/unit -m "not slow" on the locked environment: 8117 passed, 0 failed.
• test_openai_gpt_adapter, test_forward_padding_side, test_left_padding_positions, test_generate_attention_mask, test_inspect_attention_mask, test_encoder_right_padding_mask and test_attention_score_sentinel: 118 passed.
• test_mask_slicing, test_openai_gpt_adapter and test_forward_padding_side also pass on transformers 5.20.2 / torch 2.14.1.
• black, isort, pycln and mypy . are clean.

I did not run the full integration or acceptance suites locally.

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility