T5Attention support for cross-attention (#2654)

* fix AttnProcessor2_0

Fix use of AttnProcessor2_0 for cross attention with mask

* added scale_qk and out_bias flags

* fixed for xformers

* check if it has scale argument

* Update cross_attention.py

* check torch version

* fix sliced attn

* style

* set scale

* fix test

* fixed addedKV processor

* revert back AttnProcessor2_0

* if missing if

* fix inner_dim

---------
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

src/diffusers/models/cross_attention.py

@@ -59,6 +59,8 @@ class CrossAttention(nn.Module):
         cross_attention_norm: bool = False,
         added_kv_proj_dim: Optional[int] = None,
         norm_num_groups: Optional[int] = None,
+        out_bias: bool = True,
+        scale_qk: bool = True,
         processor: Optional["AttnProcessor"] = None,
     ):
         super().__init__()
@@ -68,7 +70,7 @@ class CrossAttention(nn.Module):
         self.upcast_softmax = upcast_softmax
         self.cross_attention_norm = cross_attention_norm
 
-        self.scale = dim_head**-0.5
+        self.scale = dim_head**-0.5 if scale_qk else 1.0
 
         self.heads = heads
         # for slice_size > 0 the attention score computation
@@ -95,14 +97,17 @@ class CrossAttention(nn.Module):
             self.add_v_proj = nn.Linear(added_kv_proj_dim, cross_attention_dim)
 
         self.to_out = nn.ModuleList([])
-        self.to_out.append(nn.Linear(inner_dim, query_dim))
+        self.to_out.append(nn.Linear(inner_dim, query_dim, bias=out_bias))
         self.to_out.append(nn.Dropout(dropout))
 
         # set attention processor
-        # We use the AttnProcessor2_0 by default when torch2.x is used which uses
+        # We use the AttnProcessor2_0 by default when torch 2.x is used which uses
         # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention
+        # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1
         if processor is None:
-            processor = AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else CrossAttnProcessor()
+            processor = (
+                AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and scale_qk else CrossAttnProcessor()
+            )
         self.set_processor(processor)
 
     def set_use_memory_efficient_attention_xformers(
@@ -295,7 +300,9 @@ class CrossAttnProcessor:
         encoder_hidden_states=None,
         attention_mask=None,
     ):
-        batch_size, sequence_length, _ = hidden_states.shape
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
         attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
         query = attn.to_q(hidden_states)
 
@@ -362,7 +369,9 @@ class LoRACrossAttnProcessor(nn.Module):
     def __call__(
         self, attn: CrossAttention, hidden_states, encoder_hidden_states=None, attention_mask=None, scale=1.0
     ):
-        batch_size, sequence_length, _ = hidden_states.shape
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
         attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
 
         query = attn.to_q(hidden_states) + scale * self.to_q_lora(hidden_states)
@@ -435,7 +444,9 @@ class XFormersCrossAttnProcessor:
         self.attention_op = attention_op
 
     def __call__(self, attn: CrossAttention, hidden_states, encoder_hidden_states=None, attention_mask=None):
-        batch_size, sequence_length, _ = hidden_states.shape
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
 
         attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
 
@@ -454,7 +465,7 @@ class XFormersCrossAttnProcessor:
         value = attn.head_to_batch_dim(value).contiguous()
 
         hidden_states = xformers.ops.memory_efficient_attention(
-            query, key, value, attn_bias=attention_mask, op=self.attention_op
+            query, key, value, attn_bias=attention_mask, op=self.attention_op, scale=attn.scale
         )
         hidden_states = hidden_states.to(query.dtype)
         hidden_states = attn.batch_to_head_dim(hidden_states)
@@ -472,7 +483,10 @@ class AttnProcessor2_0:
             raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
 
     def __call__(self, attn: CrossAttention, hidden_states, encoder_hidden_states=None, attention_mask=None):
-        batch_size, sequence_length, inner_dim = hidden_states.shape
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        inner_dim = hidden_states.shape[-1]
 
         if attention_mask is not None:
             attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
@@ -496,6 +510,7 @@ class AttnProcessor2_0:
         value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
 
         # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
         hidden_states = F.scaled_dot_product_attention(
             query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
         )
@@ -527,7 +542,9 @@ class LoRAXFormersCrossAttnProcessor(nn.Module):
     def __call__(
         self, attn: CrossAttention, hidden_states, encoder_hidden_states=None, attention_mask=None, scale=1.0
     ):
-        batch_size, sequence_length, _ = hidden_states.shape
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
         attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
 
         query = attn.to_q(hidden_states) + scale * self.to_q_lora(hidden_states)
@@ -542,7 +559,7 @@ class LoRAXFormersCrossAttnProcessor(nn.Module):
         value = attn.head_to_batch_dim(value).contiguous()
 
         hidden_states = xformers.ops.memory_efficient_attention(
-            query, key, value, attn_bias=attention_mask, op=self.attention_op
+            query, key, value, attn_bias=attention_mask, op=self.attention_op, scale=attn.scale
         )
         hidden_states = attn.batch_to_head_dim(hidden_states)
 
@@ -559,8 +576,9 @@ class SlicedAttnProcessor:
         self.slice_size = slice_size
 
     def __call__(self, attn: CrossAttention, hidden_states, encoder_hidden_states=None, attention_mask=None):
-        batch_size, sequence_length, _ = hidden_states.shape
-
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
         attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
 
         query = attn.to_q(hidden_states)
@@ -577,12 +595,12 @@ class SlicedAttnProcessor:
         key = attn.head_to_batch_dim(key)
         value = attn.head_to_batch_dim(value)
 
-        batch_size_attention = query.shape[0]
+        batch_size_attention, query_tokens, _ = query.shape
         hidden_states = torch.zeros(
-            (batch_size_attention, sequence_length, dim // attn.heads), device=query.device, dtype=query.dtype
+            (batch_size_attention, query_tokens, dim // attn.heads), device=query.device, dtype=query.dtype
         )
 
-        for i in range(hidden_states.shape[0] // self.slice_size):
+        for i in range(batch_size_attention // self.slice_size):
             start_idx = i * self.slice_size
             end_idx = (i + 1) * self.slice_size
 
@@ -638,12 +656,12 @@ class SlicedAttnAddedKVProcessor:
         key = torch.cat([encoder_hidden_states_key_proj, key], dim=1)
         value = torch.cat([encoder_hidden_states_value_proj, value], dim=1)
 
-        batch_size_attention = query.shape[0]
+        batch_size_attention, query_tokens, _ = query.shape
         hidden_states = torch.zeros(
-            (batch_size_attention, sequence_length, dim // attn.heads), device=query.device, dtype=query.dtype
+            (batch_size_attention, query_tokens, dim // attn.heads), device=query.device, dtype=query.dtype
         )
 
-        for i in range(hidden_states.shape[0] // self.slice_size):
+        for i in range(batch_size_attention // self.slice_size):
             start_idx = i * self.slice_size
             end_idx = (i + 1) * self.slice_size
 

tests/models/test_models_unet_2d_condition.py

@@ -118,7 +118,8 @@ class UNet2DConditionModelTests(ModelTesterMixin, unittest.TestCase):
         model.enable_xformers_memory_efficient_attention()
 
         assert (
-            model.mid_block.attentions[0].transformer_blocks[0].attn1._use_memory_efficient_attention_xformers
+            model.mid_block.attentions[0].transformer_blocks[0].attn1.processor.__class__.__name__
+            == "XFormersCrossAttnProcessor"
         ), "xformers is not enabled"
 
     @unittest.skipIf(torch_device == "mps", "Gradient checkpointing skipped on MPS")