[YHC66]

Description

MLAAttentionBridge.forward (DeepSeek-V2/V3, GLM-4.7-Flash) recomputes attention by hand, but it only applies a mask when one is passed, and only as an additive float mask:

if attention_mask is not None:
    attention_mask = self._normalize_compatibility_mask_sentinel(attention_mask)
    attn_scores = attn_scores + attention_mask

When the wrapped HF model runs with attn_implementation="sdpa" (the transformers default), HF does not build a float 4D mask for the attention modules:

• with no padding it passes attention_mask=None and relies on SDPA's is_causal, so the bridge attends bidirectionally;
• with padding it passes a boolean 4D mask (True = attend), which the bridge adds to the scores as 0/1, so neither the causal mask nor the padding is applied.

boot_transformers forces eager attention, so it is not affected. build_bridge_from_module leaves the supplied model alone (as documented), so a DeepSeek model loaded with default settings and wrapped this way gets wrong logits, and earlier positions depend on later tokens. With a tiny random DeepseekV3ForCausalLM on sdpa (dev 2ad49297, transformers 5.20.1):

	before	after
bridge vs HF log-probs, no padding (V3 / V2)	1.0e-1 / 1.6e-1	4.8e-7 / 4.8e-7
change in positions 0–2 when only tokens 3–5 change	6.1e-2	0
bridge vs HF on the real tokens of a left-padded batch	1.0e-1	4.8e-7

Eager models were already correct and are unchanged.

The fix replaces those three lines with AttentionBridge._apply_reconstruct_attention_mask, which the other reconstructing bridges (PositionEmbeddingsAttentionBridge, JointQKVAttentionBridge, ...) already use: a 4D mask is used as given (bool or float), otherwise the causal mask is built, with the KV-cache offset handled. The compatibility-mode -inf sentinel handling is the same as before. GlmMoeDsaAttentionBridge has its own forward and forces eager, so it is not touched.

Tests (test_mla_attention_bridge.py):

• test_output_matches_hf already compared the bridge against the sdpa HF attention module with attention_mask=None, but with max_diff < 0.15, which was loose enough to hide the missing causal mask. It is now 1e-5 (the actual difference after the fix is around 1e-7).
• New test_boolean_padding_mask_matches_hf passes a left-padded boolean 4D mask to both and compares the non-padded positions.

Both fail on dev and pass with the fix.

Type of change

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

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

Local runs: tests/unit on the locked environment (transformers 5.13.0, torch 2.11.0): 7676 passed, 0 failed; the MLA/DeepSeek/GLM-4.7-Flash/GLM-MoE-DSA unit and integration tests on both 5.13.0 and 5.20.1 (126 passed each), plus black, isort, pycln and mypy ..