Improve memory text to video (huggingface#3930)

* Improve memory text to video

* Apply suggestions from code review

* add test

* Apply suggestions from code review

Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

* finish test setup

---------

Co-authored-by: Pedro Cuenca <pedro@huggingface.co>

Author: patrickvonplaten

models/attention.py

@@ -119,6 +119,15 @@ def __init__(
         self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
         self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
 
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
     def forward(
         self,
         hidden_states: torch.FloatTensor,
@@ -141,6 +150,7 @@ def forward(
             norm_hidden_states = self.norm1(hidden_states)
 
         cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+
         attn_output = self.attn1(
             norm_hidden_states,
             encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
@@ -171,7 +181,20 @@ def forward(
         if self.use_ada_layer_norm_zero:
             norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
 
-        ff_output = self.ff(norm_hidden_states)
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
+                raise ValueError(
+                    f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+                )
+
+            num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
+            ff_output = torch.cat(
+                [self.ff(hid_slice) for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)],
+                dim=self._chunk_dim,
+            )
+        else:
+            ff_output = self.ff(norm_hidden_states)
 
         if self.use_ada_layer_norm_zero:
             ff_output = gate_mlp.unsqueeze(1) * ff_output

models/unet_3d_condition.py

@@ -389,6 +389,46 @@ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
         for name, module in self.named_children():
             fn_recursive_attn_processor(name, module, processor)
 
+    def enable_forward_chunking(self, chunk_size=None, dim=0):
+        """
+        Sets the attention processor to use [feed forward
+        chunking](https://huggingface.co/blog/reformer#2-chunked-feed-forward-layers).
+
+        Parameters:
+            chunk_size (`int`, *optional*):
+                The chunk size of the feed-forward layers. If not specified, will run feed-forward layer individually
+                over each tensor of dim=`dim`.
+            dim (`int`, *optional*, defaults to `0`):
+                The dimension over which the feed-forward computation should be chunked. Choose between dim=0 (batch)
+                or dim=1 (sequence length).
+        """
+        if dim not in [0, 1]:
+            raise ValueError(f"Make sure to set `dim` to either 0 or 1, not {dim}")
+
+        # By default chunk size is 1
+        chunk_size = chunk_size or 1
+
+        def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
+            if hasattr(module, "set_chunk_feed_forward"):
+                module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
+
+            for child in module.children():
+                fn_recursive_feed_forward(child, chunk_size, dim)
+
+        for module in self.children():
+            fn_recursive_feed_forward(module, chunk_size, dim)
+
+    def disable_forward_chunking(self):
+        def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
+            if hasattr(module, "set_chunk_feed_forward"):
+                module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
+
+            for child in module.children():
+                fn_recursive_feed_forward(child, chunk_size, dim)
+
+        for module in self.children():
+            fn_recursive_feed_forward(module, None, 0)
+
     # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
     def set_default_attn_processor(self):
         """

pipelines/text_to_video_synthesis/pipeline_text_to_video_synth.py

@@ -634,6 +634,9 @@ def __call__(
         # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
+        # 6.1 Chunk feed-forward computation to save memory
+        self.unet.enable_forward_chunking(chunk_size=1, dim=1)
+
         # 7. Denoising loop
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
         with self.progress_bar(total=num_inference_steps) as progress_bar:

pipelines/text_to_video_synthesis/pipeline_text_to_video_synth_img2img.py

@@ -709,6 +709,9 @@ def __call__(
         # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
+        # 6.1 Chunk feed-forward computation to save memory
+        self.unet.enable_forward_chunking(chunk_size=1, dim=1)
+
         # 7. Denoising loop
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
         with self.progress_bar(total=num_inference_steps) as progress_bar: