[refactor] make set_attention_slice recursive (huggingface#1532)

* make attn slice recursive

* remove set_attention_slice from blocks

* fix copies

* make enable_attention_slicing base class method of DiffusionPipeline

* fix set_attention_slice

* fix set_attention_slice

* fix copies

* add tests

* up

* up

* up

* update

* up

* uP

Co-authored-by: Patrick von Platen <[email protected]>

Author: patil-suraj

src/diffusers/models/attention.py

@@ -174,10 +174,6 @@ def __init__(
             self.norm_out = nn.LayerNorm(inner_dim)
             self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1)
 
-    def _set_attention_slice(self, slice_size):
-        for block in self.transformer_blocks:
-            block._set_attention_slice(slice_size)
-
     def forward(self, hidden_states, encoder_hidden_states=None, timestep=None, return_dict: bool = True):
         """
         Args:
@@ -448,10 +444,6 @@ def __init__(
                     f" correctly and a GPU is available: {e}"
                 )
 
-    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")
@@ -534,6 +526,7 @@ def __init__(
         # for slice_size > 0 the attention score computation
         # is split across the batch axis to save memory
         # You can set slice_size with `set_attention_slice`
+        self.sliceable_head_dim = heads
         self._slice_size = None
         self._use_memory_efficient_attention_xformers = False
 
@@ -559,6 +552,12 @@ def reshape_batch_dim_to_heads(self, tensor):
         tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
         return tensor
 
+    def set_attention_slice(self, slice_size):
+        if slice_size is not None and slice_size > self.sliceable_head_dim:
+            raise ValueError(f"slice_size {slice_size} has to be smaller or equal to {self.sliceable_head_dim}.")
+
+        self._slice_size = slice_size
+
     def forward(self, hidden_states, context=None, mask=None):
         batch_size, sequence_length, _ = hidden_states.shape
 

src/diffusers/models/unet_2d_blocks.py

@@ -401,23 +401,6 @@ def __init__(
         self.attentions = nn.ModuleList(attentions)
         self.resnets = nn.ModuleList(resnets)
 
-    def set_attention_slice(self, slice_size):
-        head_dims = self.attn_num_head_channels
-        head_dims = [head_dims] if isinstance(head_dims, int) else head_dims
-        if slice_size is not None and any(dim % slice_size != 0 for dim in head_dims):
-            raise ValueError(
-                f"Make sure slice_size {slice_size} is a common divisor of "
-                f"the number of heads used in cross_attention: {head_dims}"
-            )
-        if slice_size is not None and slice_size > min(head_dims):
-            raise ValueError(
-                f"slice_size {slice_size} has to be smaller or equal to "
-                f"the lowest number of heads used in cross_attention: min({head_dims}) = {min(head_dims)}"
-            )
-
-        for attn in self.attentions:
-            attn._set_attention_slice(slice_size)
-
     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:]):
@@ -595,23 +578,6 @@ def __init__(
 
         self.gradient_checkpointing = False
 
-    def set_attention_slice(self, slice_size):
-        head_dims = self.attn_num_head_channels
-        head_dims = [head_dims] if isinstance(head_dims, int) else head_dims
-        if slice_size is not None and any(dim % slice_size != 0 for dim in head_dims):
-            raise ValueError(
-                f"Make sure slice_size {slice_size} is a common divisor of "
-                f"the number of heads used in cross_attention: {head_dims}"
-            )
-        if slice_size is not None and slice_size > min(head_dims):
-            raise ValueError(
-                f"slice_size {slice_size} has to be smaller or equal to "
-                f"the lowest number of heads used in cross_attention: min({head_dims}) = {min(head_dims)}"
-            )
-
-        for attn in self.attentions:
-            attn._set_attention_slice(slice_size)
-
     def forward(self, hidden_states, temb=None, encoder_hidden_states=None):
         output_states = ()
 
@@ -1190,25 +1156,6 @@ def __init__(
 
         self.gradient_checkpointing = False
 
-    def set_attention_slice(self, slice_size):
-        head_dims = self.attn_num_head_channels
-        head_dims = [head_dims] if isinstance(head_dims, int) else head_dims
-        if slice_size is not None and any(dim % slice_size != 0 for dim in head_dims):
-            raise ValueError(
-                f"Make sure slice_size {slice_size} is a common divisor of "
-                f"the number of heads used in cross_attention: {head_dims}"
-            )
-        if slice_size is not None and slice_size > min(head_dims):
-            raise ValueError(
-                f"slice_size {slice_size} has to be smaller or equal to "
-                f"the lowest number of heads used in cross_attention: min({head_dims}) = {min(head_dims)}"
-            )
-
-        for attn in self.attentions:
-            attn._set_attention_slice(slice_size)
-
-        self.gradient_checkpointing = False
-
     def forward(
         self,
         hidden_states,

src/diffusers/models/unet_2d_condition.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from dataclasses import dataclass
-from typing import Optional, Tuple, Union
+from typing import List, Optional, Tuple, Union
 
 import torch
 import torch.nn as nn
@@ -229,28 +229,69 @@ def __init__(
         self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, kernel_size=3, padding=1)
 
     def set_attention_slice(self, slice_size):
-        head_dims = self.config.attention_head_dim
-        head_dims = [head_dims] if isinstance(head_dims, int) else head_dims
-        if slice_size is not None and any(dim % slice_size != 0 for dim in head_dims):
-            raise ValueError(
-                f"Make sure slice_size {slice_size} is a common divisor of "
-                f"the number of heads used in cross_attention: {head_dims}"
-            )
-        if slice_size is not None and slice_size > min(head_dims):
-            raise ValueError(
-                f"slice_size {slice_size} has to be smaller or equal to "
-                f"the lowest number of heads used in cross_attention: min({head_dims}) = {min(head_dims)}"
-            )
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                `"max"`, maxium amount of memory will be saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_slicable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_slicable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_slicable_dims(module)
 
-        for block in self.down_blocks:
-            if hasattr(block, "attentions") and block.attentions is not None:
-                block.set_attention_slice(slice_size)
+        num_slicable_layers = len(sliceable_head_dims)
 
-        self.mid_block.set_attention_slice(slice_size)
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_slicable_layers * [1]
+
+        slice_size = num_slicable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
 
-        for block in self.up_blocks:
-            if hasattr(block, "attentions") and block.attentions is not None:
-                block.set_attention_slice(slice_size)
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
 
     def _set_gradient_checkpointing(self, module, value=False):
         if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D, CrossAttnUpBlock2D, UpBlock2D)):

src/diffusers/pipeline_utils.py

@@ -839,3 +839,34 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
             module = getattr(self, module_name)
             if isinstance(module, torch.nn.Module):
                 fn_recursive_set_mem_eff(module)
+
+    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                `"max"`, maxium amount of memory will be saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        self.set_attention_slice(slice_size)
+
+    def disable_attention_slicing(self):
+        r"""
+        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
+        back to computing attention in one step.
+        """
+        # set slice_size = `None` to disable `attention slicing`
+        self.enable_attention_slicing(None)
+
+    def set_attention_slice(self, slice_size: Optional[int]):
+        module_names, _, _ = self.extract_init_dict(dict(self.config))
+        for module_name in module_names:
+            module = getattr(self, module_name)
+            if isinstance(module, torch.nn.Module) and hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size)

src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion.py

@@ -166,38 +166,6 @@ def __init__(
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
-    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
-        r"""
-        Enable sliced attention computation.
-
-        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
-        in several steps. This is useful to save some memory in exchange for a small speed decrease.
-
-        Args:
-            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
-                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
-                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
-                `attention_head_dim` must be a multiple of `slice_size`.
-        """
-        if slice_size == "auto":
-            if isinstance(self.unet.config.attention_head_dim, int):
-                # half the attention head size is usually a good trade-off between
-                # speed and memory
-                slice_size = self.unet.config.attention_head_dim // 2
-            else:
-                # if `attention_head_dim` is a list, take the smallest head size
-                slice_size = min(self.unet.config.attention_head_dim)
-
-        self.unet.set_attention_slice(slice_size)
-
-    def disable_attention_slicing(self):
-        r"""
-        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
-        back to computing attention in one step.
-        """
-        # set slice_size = `None` to disable `attention slicing`
-        self.enable_attention_slicing(None)
-
     def enable_vae_slicing(self):
         r"""
         Enable sliced VAE decoding.

src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py

@@ -179,38 +179,6 @@ def __init__(
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
-    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
-        r"""
-        Enable sliced attention computation.
-
-        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
-        in several steps. This is useful to save some memory in exchange for a small speed decrease.
-
-        Args:
-            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
-                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
-                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
-                `attention_head_dim` must be a multiple of `slice_size`.
-        """
-        if slice_size == "auto":
-            if isinstance(self.unet.config.attention_head_dim, int):
-                # half the attention head size is usually a good trade-off between
-                # speed and memory
-                slice_size = self.unet.config.attention_head_dim // 2
-            else:
-                # if `attention_head_dim` is a list, take the smallest head size
-                slice_size = min(self.unet.config.attention_head_dim)
-
-        self.unet.set_attention_slice(slice_size)
-
-    def disable_attention_slicing(self):
-        r"""
-        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
-        back to computing attention in one step.
-        """
-        # set slice_size = `None` to disable `attention slicing`
-        self.enable_attention_slicing(None)
-
     def enable_sequential_cpu_offload(self, gpu_id=0):
         r"""
         Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,

src/diffusers/pipelines/stable_diffusion/pipeline_cycle_diffusion.py

@@ -209,40 +209,6 @@ def __init__(
         )
         self.register_to_config(requires_safety_checker=requires_safety_checker)
 
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_attention_slicing
-    def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
-        r"""
-        Enable sliced attention computation.
-
-        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
-        in several steps. This is useful to save some memory in exchange for a small speed decrease.
-
-        Args:
-            slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
-                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
-                a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
-                `attention_head_dim` must be a multiple of `slice_size`.
-        """
-        if slice_size == "auto":
-            if isinstance(self.unet.config.attention_head_dim, int):
-                # half the attention head size is usually a good trade-off between
-                # speed and memory
-                slice_size = self.unet.config.attention_head_dim // 2
-            else:
-                # if `attention_head_dim` is a list, take the smallest head size
-                slice_size = min(self.unet.config.attention_head_dim)
-
-        self.unet.set_attention_slice(slice_size)
-
-    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_attention_slicing
-    def disable_attention_slicing(self):
-        r"""
-        Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
-        back to computing attention in one step.
-        """
-        # set slice_size = `None` to disable `attention slicing`
-        self.enable_attention_slicing(None)
-
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_sequential_cpu_offload
     def enable_sequential_cpu_offload(self, gpu_id=0):
         r"""