Attn added kv processor torch 2.0 block (huggingface#3023)

add AttnAddedKVProcessor2_0 block

Author: williamberman

src/diffusers/models/attention_processor.py

@@ -255,11 +255,15 @@ def batch_to_head_dim(self, tensor):
         tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
         return tensor
 
-    def head_to_batch_dim(self, tensor):
+    def head_to_batch_dim(self, tensor, out_dim=3):
         head_size = self.heads
         batch_size, seq_len, dim = tensor.shape
         tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size)
-        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size * head_size, seq_len, dim // head_size)
+        tensor = tensor.permute(0, 2, 1, 3)
+
+        if out_dim == 3:
+            tensor = tensor.reshape(batch_size * head_size, seq_len, dim // head_size)
+
         return tensor
 
     def get_attention_scores(self, query, key, attention_mask=None):
@@ -293,7 +297,7 @@ def get_attention_scores(self, query, key, attention_mask=None):
 
         return attention_probs
 
-    def prepare_attention_mask(self, attention_mask, target_length, batch_size=None):
+    def prepare_attention_mask(self, attention_mask, target_length, batch_size=None, out_dim=3):
         if batch_size is None:
             deprecate(
                 "batch_size=None",
@@ -320,8 +324,13 @@ def prepare_attention_mask(self, attention_mask, target_length, batch_size=None)
             else:
                 attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)
 
-        if attention_mask.shape[0] < batch_size * head_size:
-            attention_mask = attention_mask.repeat_interleave(head_size, dim=0)
+        if out_dim == 3:
+            if attention_mask.shape[0] < batch_size * head_size:
+                attention_mask = attention_mask.repeat_interleave(head_size, dim=0)
+        elif out_dim == 4:
+            attention_mask = attention_mask.unsqueeze(1)
+            attention_mask = attention_mask.repeat_interleave(head_size, dim=1)
+
         return attention_mask
 
     def norm_encoder_hidden_states(self, encoder_hidden_states):
@@ -499,6 +508,64 @@ def __call__(self, attn: Attention, hidden_states, encoder_hidden_states=None, a
         return hidden_states
 
 
+class AttnAddedKVProcessor2_0:
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "AttnAddedKVProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+            )
+
+    def __call__(self, attn: Attention, hidden_states, encoder_hidden_states=None, attention_mask=None):
+        residual = hidden_states
+        hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2)
+        batch_size, sequence_length, _ = hidden_states.shape
+
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size, out_dim=4)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+        query = attn.head_to_batch_dim(query, out_dim=4)
+
+        encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+        encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+        encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj, out_dim=4)
+        encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj, out_dim=4)
+
+        if not attn.only_cross_attention:
+            key = attn.to_k(hidden_states)
+            value = attn.to_v(hidden_states)
+            key = attn.head_to_batch_dim(key, out_dim=4)
+            value = attn.head_to_batch_dim(value, out_dim=4)
+            key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
+            value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
+        else:
+            key = encoder_hidden_states_key_proj
+            value = encoder_hidden_states_value_proj
+
+        # 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
+        )
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, residual.shape[1])
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape)
+        hidden_states = hidden_states + residual
+
+        return hidden_states
+
+
 class XFormersAttnProcessor:
     def __init__(self, attention_op: Optional[Callable] = None):
         self.attention_op = attention_op
@@ -764,6 +831,7 @@ def __call__(self, attn: "Attention", hidden_states, encoder_hidden_states=None,
     SlicedAttnProcessor,
     AttnAddedKVProcessor,
     SlicedAttnAddedKVProcessor,
+    AttnAddedKVProcessor2_0,
     LoRAAttnProcessor,
     LoRAXFormersAttnProcessor,
 ]

src/diffusers/models/unet_2d_blocks.py

@@ -15,10 +15,11 @@
 
 import numpy as np
 import torch
+import torch.nn.functional as F
 from torch import nn
 
 from .attention import AdaGroupNorm, AttentionBlock
-from .attention_processor import Attention, AttnAddedKVProcessor
+from .attention_processor import Attention, AttnAddedKVProcessor, AttnAddedKVProcessor2_0
 from .dual_transformer_2d import DualTransformer2DModel
 from .resnet import Downsample2D, FirDownsample2D, FirUpsample2D, KDownsample2D, KUpsample2D, ResnetBlock2D, Upsample2D
 from .transformer_2d import Transformer2DModel
@@ -612,6 +613,10 @@ def __init__(
         attentions = []
 
         for _ in range(num_layers):
+            processor = (
+                AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
+            )
+
             attentions.append(
                 Attention(
                     query_dim=in_channels,
@@ -624,7 +629,7 @@ def __init__(
                     upcast_softmax=True,
                     only_cross_attention=only_cross_attention,
                     cross_attention_norm=cross_attention_norm,
-                    processor=AttnAddedKVProcessor(),
+                    processor=processor,
                 )
             )
             resnets.append(
@@ -1396,6 +1401,11 @@ def __init__(
                     skip_time_act=skip_time_act,
                 )
             )
+
+            processor = (
+                AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
+            )
+
             attentions.append(
                 Attention(
                     query_dim=out_channels,
@@ -1408,7 +1418,7 @@ def __init__(
                     upcast_softmax=True,
                     only_cross_attention=only_cross_attention,
                     cross_attention_norm=cross_attention_norm,
-                    processor=AttnAddedKVProcessor(),
+                    processor=processor,
                 )
             )
         self.attentions = nn.ModuleList(attentions)
@@ -2399,6 +2409,11 @@ def __init__(
                     skip_time_act=skip_time_act,
                 )
             )
+
+            processor = (
+                AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
+            )
+
             attentions.append(
                 Attention(
                     query_dim=out_channels,
@@ -2411,7 +2426,7 @@ def __init__(
                     upcast_softmax=True,
                     only_cross_attention=only_cross_attention,
                     cross_attention_norm=cross_attention_norm,
-                    processor=AttnAddedKVProcessor(),
+                    processor=processor,
                 )
             )
         self.attentions = nn.ModuleList(attentions)

src/diffusers/pipelines/versatile_diffusion/modeling_text_unet.py

@@ -8,7 +8,12 @@
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...models import ModelMixin
 from ...models.attention import Attention
-from ...models.attention_processor import AttentionProcessor, AttnAddedKVProcessor, AttnProcessor
+from ...models.attention_processor import (
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnAddedKVProcessor2_0,
+    AttnProcessor,
+)
 from ...models.dual_transformer_2d import DualTransformer2DModel
 from ...models.embeddings import GaussianFourierProjection, TimestepEmbedding, Timesteps
 from ...models.transformer_2d import Transformer2DModel
@@ -1545,6 +1550,10 @@ def __init__(
         attentions = []
 
         for _ in range(num_layers):
+            processor = (
+                AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
+            )
+
             attentions.append(
                 Attention(
                     query_dim=in_channels,
@@ -1557,7 +1566,7 @@ def __init__(
                     upcast_softmax=True,
                     only_cross_attention=only_cross_attention,
                     cross_attention_norm=cross_attention_norm,
-                    processor=AttnAddedKVProcessor(),
+                    processor=processor,
                 )
             )
             resnets.append(

tests/pipelines/unclip/test_unclip_image_variation.py

@@ -421,7 +421,12 @@ class DummyScheduler:
     def test_attention_slicing_forward_pass(self):
         test_max_difference = torch_device == "cpu"
 
-        self._test_attention_slicing_forward_pass(test_max_difference=test_max_difference)
+        # Check is relaxed because there is not a torch 2.0 sliced attention added kv processor
+        expected_max_diff = 1e-2
+
+        self._test_attention_slicing_forward_pass(
+            test_max_difference=test_max_difference, expected_max_diff=expected_max_diff
+        )
 
     # Overriding PipelineTesterMixin::test_inference_batch_single_identical
     # because UnCLIP undeterminism requires a looser check.