Enabling gradient checkpointing for VAE (#2536)

* updated black format

* update black format

* make style format

* updated line endings

* update code formatting

* Update examples/research_projects/onnxruntime/text_to_image/train_text_to_image.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/diffusers/models/vae.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/diffusers/models/vae.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* added vae gradient checkpointing test

* make style

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

examples/research_projects/onnxruntime/text_to_image/train_text_to_image.py

@@ -412,6 +412,7 @@ def main():
 
     if args.gradient_checkpointing:
         unet.enable_gradient_checkpointing()
+        vae.enable_gradient_checkpointing()
 
     # Enable TF32 for faster training on Ampere GPUs,
     # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices

src/diffusers/models/autoencoder_kl.py

@@ -65,6 +65,8 @@ class AutoencoderKL(ModelMixin, ConfigMixin):
             Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
     """
 
+    _supports_gradient_checkpointing = True
+
     @register_to_config
     def __init__(
         self,
@@ -121,6 +123,10 @@ class AutoencoderKL(ModelMixin, ConfigMixin):
         self.tile_latent_min_size = int(sample_size / (2 ** (len(self.block_out_channels) - 1)))
         self.tile_overlap_factor = 0.25
 
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, (Encoder, Decoder)):
+            module.gradient_checkpointing = value
+
     def enable_tiling(self, use_tiling: bool = True):
         r"""
         Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to

src/diffusers/models/cross_attention.py

@@ -24,9 +24,7 @@ from .attention_processor import (  # noqa: F401
     SlicedAttnProcessor,
     XFormersAttnProcessor,
 )
-from .attention_processor import (  # noqa: F401
-    AttnProcessor as AttnProcessorRename,
-)
+from .attention_processor import AttnProcessor as AttnProcessorRename  # noqa: F401
 
 
 deprecate(

src/diffusers/models/vae.py

@@ -50,7 +50,13 @@ class Encoder(nn.Module):
         super().__init__()
         self.layers_per_block = layers_per_block
 
-        self.conv_in = torch.nn.Conv2d(in_channels, block_out_channels[0], kernel_size=3, stride=1, padding=1)
+        self.conv_in = torch.nn.Conv2d(
+            in_channels,
+            block_out_channels[0],
+            kernel_size=3,
+            stride=1,
+            padding=1,
+        )
 
         self.mid_block = None
         self.down_blocks = nn.ModuleList([])
@@ -96,10 +102,28 @@ class Encoder(nn.Module):
         conv_out_channels = 2 * out_channels if double_z else out_channels
         self.conv_out = nn.Conv2d(block_out_channels[-1], conv_out_channels, 3, padding=1)
 
+        self.gradient_checkpointing = False
+
     def forward(self, x):
         sample = x
         sample = self.conv_in(sample)
 
+        if self.training and self.gradient_checkpointing:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    return module(*inputs)
+
+                return custom_forward
+
+            # down
+            for down_block in self.down_blocks:
+                sample = torch.utils.checkpoint.checkpoint(create_custom_forward(down_block), sample)
+
+            # middle
+            sample = torch.utils.checkpoint.checkpoint(create_custom_forward(self.mid_block), sample)
+
+        else:
             # down
             for down_block in self.down_blocks:
                 sample = down_block(sample)
@@ -129,7 +153,13 @@ class Decoder(nn.Module):
         super().__init__()
         self.layers_per_block = layers_per_block
 
-        self.conv_in = nn.Conv2d(in_channels, block_out_channels[-1], kernel_size=3, stride=1, padding=1)
+        self.conv_in = nn.Conv2d(
+            in_channels,
+            block_out_channels[-1],
+            kernel_size=3,
+            stride=1,
+            padding=1,
+        )
 
         self.mid_block = None
         self.up_blocks = nn.ModuleList([])
@@ -176,10 +206,27 @@ class Decoder(nn.Module):
         self.conv_act = nn.SiLU()
         self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, 3, padding=1)
 
+        self.gradient_checkpointing = False
+
     def forward(self, z):
         sample = z
         sample = self.conv_in(sample)
 
+        if self.training and self.gradient_checkpointing:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    return module(*inputs)
+
+                return custom_forward
+
+            # middle
+            sample = torch.utils.checkpoint.checkpoint(create_custom_forward(self.mid_block), sample)
+
+            # up
+            for up_block in self.up_blocks:
+                sample = torch.utils.checkpoint.checkpoint(create_custom_forward(up_block), sample)
+        else:
             # middle
             sample = self.mid_block(sample)
 

tests/models/test_models_vae.py

@@ -68,6 +68,47 @@ class AutoencoderKLTests(ModelTesterMixin, unittest.TestCase):
     def test_training(self):
         pass
 
+    @unittest.skipIf(torch_device == "mps", "Gradient checkpointing skipped on MPS")
+    def test_gradient_checkpointing(self):
+        # enable deterministic behavior for gradient checkpointing
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+        model = self.model_class(**init_dict)
+        model.to(torch_device)
+
+        assert not model.is_gradient_checkpointing and model.training
+
+        out = model(**inputs_dict).sample
+        # run the backwards pass on the model. For backwards pass, for simplicity purpose,
+        # we won't calculate the loss and rather backprop on out.sum()
+        model.zero_grad()
+
+        labels = torch.randn_like(out)
+        loss = (out - labels).mean()
+        loss.backward()
+
+        # re-instantiate the model now enabling gradient checkpointing
+        model_2 = self.model_class(**init_dict)
+        # clone model
+        model_2.load_state_dict(model.state_dict())
+        model_2.to(torch_device)
+        model_2.enable_gradient_checkpointing()
+
+        assert model_2.is_gradient_checkpointing and model_2.training
+
+        out_2 = model_2(**inputs_dict).sample
+        # run the backwards pass on the model. For backwards pass, for simplicity purpose,
+        # we won't calculate the loss and rather backprop on out.sum()
+        model_2.zero_grad()
+        loss_2 = (out_2 - labels).mean()
+        loss_2.backward()
+
+        # compare the output and parameters gradients
+        self.assertTrue((loss - loss_2).abs() < 1e-5)
+        named_params = dict(model.named_parameters())
+        named_params_2 = dict(model_2.named_parameters())
+        for name, param in named_params.items():
+            self.assertTrue(torch_all_close(param.grad.data, named_params_2[name].grad.data, atol=5e-5))
+
     def test_from_pretrained_hub(self):
         model, loading_info = AutoencoderKL.from_pretrained("fusing/autoencoder-kl-dummy", output_loading_info=True)
         self.assertIsNotNone(model)