Ignore non-causal mask in more cases with SDPA (huggingface#30138)

* update non-causal mask for sdpa

* add test

* update docstrings

* add one more test

* fix cross attention bug

* gentler atol/rtol

src/transformers/modeling_attn_mask_utils.py

@@ -413,7 +413,7 @@ def _prepare_4d_attention_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len:
     `(batch_size, key_value_length)`
 
     Args:
-        mask (`torch.Tensor` or `None`):
+        mask (`torch.Tensor`):
             A 2D attention mask of shape `(batch_size, key_value_length)`
         dtype (`torch.dtype`):
             The torch dtype the created mask shall have.
@@ -429,36 +429,25 @@ def _prepare_4d_attention_mask_for_sdpa(mask: torch.Tensor, dtype: torch.dtype,
     `(batch_size, key_value_length)`
 
     Args:
-        mask (`torch.Tensor` or `None`):
+        mask (`torch.Tensor`):
             A 2D attention mask of shape `(batch_size, key_value_length)`
         dtype (`torch.dtype`):
             The torch dtype the created mask shall have.
         tgt_len (`int`):
             The target length or query length the created mask shall have.
     """
-    batch_size, key_value_length = mask.shape
+    _, key_value_length = mask.shape
     tgt_len = tgt_len if tgt_len is not None else key_value_length
 
-    # torch.jit.trace, symbolic_trace and torchdynamo with fullgraph=True are unable to capture the controlflow `is_causal=attention_mask is None and q_len > 1`
-    # used as an SDPA argument. We keep compatibility with these tracing tools by always using SDPA's `attn_mask` argument in case we are tracing.
-    # TODO: For dynamo, rather use a check on fullgraph=True once this is possible (https://github.com/pytorch/pytorch/pull/120400).
     is_tracing = (
         torch.jit.is_tracing()
         or isinstance(mask, torch.fx.Proxy)
         or (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling())
     )
 
+    # torch.jit.trace, symbolic_trace and torchdynamo with fullgraph=True are unable to capture data-dependent controlflows.
     if not is_tracing and torch.all(mask == 1):
-        if 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)
+        return None
     else:
         return AttentionMaskConverter._expand_mask(mask=mask, dtype=dtype, tgt_len=tgt_len)
 

src/transformers/models/bert/modeling_bert.py

@@ -432,7 +432,9 @@ def forward(
         # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
         # The tgt_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create
         # a causal mask in case tgt_len == 1.
-        is_causal = True if self.is_decoder and attention_mask is None and tgt_len > 1 else False
+        is_causal = (
+            True if self.is_decoder and not is_cross_attention and attention_mask is None and tgt_len > 1 else False
+        )
 
         attn_output = torch.nn.functional.scaled_dot_product_attention(
             query_layer,

tests/models/bert/test_modeling_bert.py

@@ -16,7 +16,7 @@
 import tempfile
 import unittest
 
-from transformers import BertConfig, is_torch_available
+from transformers import AutoTokenizer, BertConfig, is_torch_available
 from transformers.models.auto import get_values
 from transformers.testing_utils import (
     CaptureLogger,
@@ -747,3 +747,36 @@ def test_inference_no_head_relative_embedding_key_query(self):
         )
 
         self.assertTrue(torch.allclose(output[:, 1:4, 1:4], expected_slice, atol=1e-4))
+
+    def test_sdpa_ignored_mask(self):
+        pkv = []
+
+        model = BertModel.from_pretrained("hf-internal-testing/tiny-random-BertModel", attn_implementation="eager")
+        model_sdpa = BertModel.from_pretrained("hf-internal-testing/tiny-random-BertModel", attn_implementation="sdpa")
+
+        model = model.eval()
+        model_sdpa = model_sdpa.eval()
+
+        for _ in range(model.config.num_hidden_layers):
+            num_heads = model.config.num_attention_heads
+            head_dim = model.config.hidden_size // model.config.num_attention_heads
+            pkv.append([torch.rand(1, num_heads, 3, head_dim), torch.rand(1, num_heads, 3, head_dim)])
+
+        tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-BertModel")
+        inp = tokenizer("I am in Paris and", return_tensors="pt")
+
+        del inp["attention_mask"]
+
+        with torch.no_grad():
+            res_eager = model(**inp)
+            res_sdpa = model_sdpa(**inp)
+            self.assertTrue(
+                torch.allclose(res_eager.last_hidden_state, res_sdpa.last_hidden_state, atol=1e-5, rtol=1e-4)
+            )
+
+            # Case where query length != kv_length.
+            res_eager = model(**inp, past_key_values=pkv)
+            res_sdpa = model_sdpa(**inp, past_key_values=pkv)
+            self.assertTrue(
+                torch.allclose(res_eager.last_hidden_state, res_sdpa.last_hidden_state, atol=1e-5, rtol=1e-4)
+            )