Up to 2x speedup on GPUs using memory efficient attention

docs/source/optimization/fp16.mdx

@@ -22,6 +22,7 @@ We present some techniques and ideas to optimize 🤗 Diffusers _inference_ for
 | fp16             | 3.61s   | x2.63   |
 | channels last    | 3.30s   | x2.88   |
 | traced UNet      | 3.21s   | x2.96   |
+| memory efficient attention  | 2.63s  | x3.61   |
 
 <em>
   obtained on NVIDIA TITAN RTX by generating a single image of size 512x512 from
@@ -290,3 +291,41 @@ pipe.unet = TracedUNet()
 with torch.inference_mode():
     image = pipe([prompt] * 1, num_inference_steps=50).images[0]
 ```
+
+
+## Memory Efficient Attention
+Recent work on optimizing the bandwitdh in the attention block have generated huge speed ups and gains in GPU memory usage. The most recent being Flash Attention (from @tridao, [code](https://github.com/HazyResearch/flash-attention), [paper](https://arxiv.org/pdf/2205.14135.pdf)) .
+Here are the speedups we obtain on a few Nvidia GPUs when running the inference at 512x512 with a batch size of 1 (one prompt):
+
+| GPU              	| Base Attention FP16 	| Memory Efficient Attention FP16 	|
+|------------------	|---------------------	|---------------------------------	|
+| NVIDIA Tesla T4  	| 3.5it/s             	| 5.5it/s                         	|
+| NVIDIA 3060 RTX  	| 4.6it/s             	| 7.8it/s                         	|
+| NVIDIA A10G      	| 8.88it/s            	| 15.6it/s                        	|
+| NVIDIA RTX A6000 	| 11.7it/s            	| 21.09it/s                       	|
+| NVIDIA TITAN RTX  | 12.51it/s         	| 18.22it/s                       	|
+| A100-SXM4-40GB    	| 18.6it/s            	| 29.it/s                        	|
+| A100-SXM-80GB    	| 18.7it/s            	| 29.5it/s                        	|
+
+To leverage it just make sure you have: 
+ - PyTorch > 1.12
+ - Cuda available
+ - Installed the [xformers](https://github.com/facebookresearch/xformers) library
+```python
+from diffusers import StableDiffusionPipeline
+import torch
+
+pipe = StableDiffusionPipeline.from_pretrained(
+    "runwayml/stable-diffusion-v1-5",
+    revision="fp16",
+    torch_dtype=torch.float16,
+).to("cuda")
+
+pipe.enable_xformers_memory_efficient_attention()
+
+with torch.inference_mode():
+    sample = pipe("a small cat")
+
+# optional: You can disable it via
+# pipe.disable_xformers_memory_efficient_attention()
+```

src/diffusers/models/attention.py

@@ -18,6 +18,15 @@
 import torch.nn.functional as F
 from torch import nn
 
+from diffusers.utils.import_utils import is_xformers_available
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
 
 class AttentionBlock(nn.Module):
     """
@@ -150,6 +159,10 @@ def _set_attention_slice(self, slice_size):
         for block in self.transformer_blocks:
             block._set_attention_slice(slice_size)
 
+    def _set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
+        for block in self.transformer_blocks:
+            block._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
+
     def forward(self, hidden_states, context=None):
         # note: if no context is given, cross-attention defaults to self-attention
         batch, channel, height, weight = hidden_states.shape
@@ -206,6 +219,32 @@ def _set_attention_slice(self, slice_size):
         self.attn1._slice_size = slice_size
         self.attn2._slice_size = slice_size
 
+    def _set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
+        if not is_xformers_available():
+            print("Here is how to install it")
+            raise ModuleNotFoundError(
+                "Refer to https://github.com/facebookresearch/xformers for more information on how to install"
+                " xformers",
+                name="xformers",
+            )
+        elif not torch.cuda.is_available():
+            raise ValueError(
+                "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is only"
+                " available for GPU "
+            )
+        else:
+            try:
+                # Make sure we can run the memory efficient attention
+                _ = xformers.ops.memory_efficient_attention(
+                    torch.randn((1, 2, 40), device="cuda"),
+                    torch.randn((1, 2, 40), device="cuda"),
+                    torch.randn((1, 2, 40), device="cuda"),
+                )
+            except Exception as e:
+                raise e
+            self.attn1._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
+            self.attn2._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
+
     def forward(self, hidden_states, context=None):
         hidden_states = self.attn1(self.norm1(hidden_states)) + hidden_states
         hidden_states = self.attn2(self.norm2(hidden_states), context=context) + hidden_states
@@ -239,6 +278,7 @@ def __init__(
         # is split across the batch axis to save memory
         # You can set slice_size with `set_attention_slice`
         self._slice_size = None
+        self._use_memory_efficient_attention_xformers = False
 
         self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
         self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
@@ -279,11 +319,13 @@ def forward(self, hidden_states, context=None, mask=None):
         # TODO(PVP) - mask is currently never used. Remember to re-implement when used
 
         # attention, what we cannot get enough of
-
-        if self._slice_size is None or query.shape[0] // self._slice_size == 1:
-            hidden_states = self._attention(query, key, value)
+        if self._use_memory_efficient_attention_xformers:
+            hidden_states = self._memory_efficient_attention_xformers(query, key, value)
         else:
-            hidden_states = self._sliced_attention(query, key, value, sequence_length, dim)
+            if self._slice_size is None or query.shape[0] // self._slice_size == 1:
+                hidden_states = self._attention(query, key, value)
+            else:
+                hidden_states = self._sliced_attention(query, key, value, sequence_length, dim)
 
         # linear proj
         hidden_states = self.to_out[0](hidden_states)
@@ -341,6 +383,11 @@ def _sliced_attention(self, query, key, value, sequence_length, dim):
         hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
         return hidden_states
 
+    def _memory_efficient_attention_xformers(self, query, key, value):
+        hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=None)
+        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
+        return hidden_states
+
 
 class FeedForward(nn.Module):
     r"""

src/diffusers/models/unet_2d_blocks.py

@@ -367,6 +367,10 @@ def set_attention_slice(self, slice_size):
         for attn in self.attentions:
             attn._set_attention_slice(slice_size)
 
+    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
+        for attn in self.attentions:
+            attn._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
+
     def forward(self, hidden_states, temb=None, encoder_hidden_states=None):
         hidden_states = self.resnets[0](hidden_states, temb)
         for attn, resnet in zip(self.attentions, self.resnets[1:]):
@@ -542,6 +546,10 @@ def set_attention_slice(self, slice_size):
         for attn in self.attentions:
             attn._set_attention_slice(slice_size)
 
+    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
+        for attn in self.attentions:
+            attn._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
+
     def forward(self, hidden_states, temb=None, encoder_hidden_states=None):
         output_states = ()
 
@@ -1117,6 +1125,10 @@ def set_attention_slice(self, slice_size):
 
         self.gradient_checkpointing = False
 
+    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
+        for attn in self.attentions:
+            attn._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
+
     def forward(
         self,
         hidden_states,

src/diffusers/models/unet_2d_condition.py

@@ -225,6 +225,17 @@ def set_attention_slice(self, slice_size):
             if hasattr(block, "attentions") and block.attentions is not None:
                 block.set_attention_slice(slice_size)
 
+    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
+        for block in self.down_blocks:
+            if hasattr(block, "attentions") and block.attentions is not None:
+                block.set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
+
+        self.mid_block.set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
+
+        for block in self.up_blocks:
+            if hasattr(block, "attentions") and block.attentions is not None:
+                block.set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)
+
     def _set_gradient_checkpointing(self, module, value=False):
         if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D, CrossAttnUpBlock2D, UpBlock2D)):
             module.gradient_checkpointing = value

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -113,6 +113,24 @@ def __init__(
             feature_extractor=feature_extractor,
         )
 
+    def enable_xformers_memory_efficient_attention(self):
+        r"""
+        Enable memory efficient attention as implemented in xformers.
+
+        When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference
+        time. Speed up at training time is not guaranteed.
+
+        Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention
+        is used.
+        """
+        self.unet.set_use_memory_efficient_attention_xformers(True)
+
+    def disable_xformers_memory_efficient_attention(self):
+        r"""
+        Disable memory efficient attention as implemented in xformers.
+        """
+        self.unet.set_use_memory_efficient_attention_xformers(False)
+
     def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
         r"""
         Enable sliced attention computation.

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py

@@ -151,6 +151,24 @@ def disable_attention_slicing(self):
         # set slice_size = `None` to disable `set_attention_slice`
         self.enable_attention_slicing(None)
 
+    def enable_xformers_memory_efficient_attention(self):
+        r"""
+        Enable memory efficient attention as implemented in xformers.
+
+        When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference
+        time. Speed up at training time is not guaranteed.
+
+        Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention
+        is used.
+        """
+        self.unet.set_use_memory_efficient_attention_xformers(True)
+
+    def disable_xformers_memory_efficient_attention(self):
+        r"""
+        Disable memory efficient attention as implemented in xformers.
+        """
+        self.unet.set_use_memory_efficient_attention_xformers(False)
+
     @torch.no_grad()
     def __call__(
         self,

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py

@@ -151,6 +151,24 @@ def disable_attention_slicing(self):
         # set slice_size = `None` to disable `attention slicing`
         self.enable_attention_slicing(None)
 
+    def enable_xformers_memory_efficient_attention(self):
+        r"""
+        Enable memory efficient attention as implemented in xformers.
+
+        When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference
+        time. Speed up at training time is not guaranteed.
+
+        Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention
+        is used.
+        """
+        self.unet.set_use_memory_efficient_attention_xformers(True)
+
+    def disable_xformers_memory_efficient_attention(self):
+        r"""
+        Disable memory efficient attention as implemented in xformers.
+        """
+        self.unet.set_use_memory_efficient_attention_xformers(False)
+
     @torch.no_grad()
     def __call__(
         self,

src/diffusers/utils/import_utils.py

@@ -168,6 +168,18 @@
 except importlib_metadata.PackageNotFoundError:
     _accelerate_available = False
 
+_xformers_available = importlib.util.find_spec("xformers") is not None
+try:
+    _xformers_version = importlib_metadata.version("xformers")
+    if _torch_available:
+        import torch
+
+        if torch.__version__ < version.Version("1.12"):
+            raise ValueError("PyTorch should be >= 1.12")
+    logger.debug(f"Successfully imported xformers version {_xformers_version}")
+except importlib_metadata.PackageNotFoundError:
+    _xformers_available = False
+
 
 def is_torch_available():
     return _torch_available
@@ -205,6 +217,10 @@ def is_scipy_available():
     return _scipy_available
 
 
+def is_xformers_available():
+    return _xformers_available
+
+
 def is_accelerate_available():
     return _accelerate_available