Enabling gradient checkpointing in eval() mode (huggingface#9878)

* refactored

Author: MikeTkachuk

examples/community/matryoshka.py

@@ -868,7 +868,7 @@ def forward(
         blocks = list(zip(self.resnets, self.attentions))
 
         for i, (resnet, attn) in enumerate(blocks):
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module, return_dict=None):
                     def custom_forward(*inputs):
@@ -1029,7 +1029,7 @@ def forward(
 
         hidden_states = self.resnets[0](hidden_states, temb)
         for attn, resnet in zip(self.attentions, self.resnets[1:]):
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module, return_dict=None):
                     def custom_forward(*inputs):
@@ -1191,7 +1191,7 @@ def forward(
 
             hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
 
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module, return_dict=None):
                     def custom_forward(*inputs):
@@ -1364,7 +1364,7 @@ def forward(
 
         # Blocks
         for block in self.transformer_blocks:
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module, return_dict=None):
                     def custom_forward(*inputs):

examples/research_projects/pixart/controlnet_pixart_alpha.py

@@ -215,7 +215,7 @@ def forward(
 
         # 2. Blocks
         for block_index, block in enumerate(self.transformer.transformer_blocks):
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
                 # rc todo: for training and gradient checkpointing
                 print("Gradient checkpointing is not supported for the controlnet transformer model, yet.")
                 exit(1)

src/diffusers/models/autoencoders/autoencoder_kl_allegro.py

@@ -506,7 +506,7 @@ def forward(self, sample: torch.Tensor) -> torch.Tensor:
         sample = self.temp_conv_in(sample)
         sample = sample + residual
 
-        if self.gradient_checkpointing:
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
 
             def create_custom_forward(module):
                 def custom_forward(*inputs):
@@ -646,7 +646,7 @@ def forward(self, sample: torch.Tensor) -> torch.Tensor:
 
         upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
 
-        if self.gradient_checkpointing:
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
 
             def create_custom_forward(module):
                 def custom_forward(*inputs):

src/diffusers/models/autoencoders/autoencoder_kl_cogvideox.py

@@ -420,7 +420,7 @@ def forward(
         for i, resnet in enumerate(self.resnets):
             conv_cache_key = f"resnet_{i}"
 
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module):
                     def create_forward(*inputs):
@@ -522,7 +522,7 @@ def forward(
         for i, resnet in enumerate(self.resnets):
             conv_cache_key = f"resnet_{i}"
 
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module):
                     def create_forward(*inputs):
@@ -636,7 +636,7 @@ def forward(
         for i, resnet in enumerate(self.resnets):
             conv_cache_key = f"resnet_{i}"
 
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module):
                     def create_forward(*inputs):
@@ -773,7 +773,7 @@ def forward(
 
         hidden_states, new_conv_cache["conv_in"] = self.conv_in(sample, conv_cache=conv_cache.get("conv_in"))
 
-        if self.training and self.gradient_checkpointing:
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
 
             def create_custom_forward(module):
                 def custom_forward(*inputs):
@@ -939,7 +939,7 @@ def forward(
 
         hidden_states, new_conv_cache["conv_in"] = self.conv_in(sample, conv_cache=conv_cache.get("conv_in"))
 
-        if self.training and self.gradient_checkpointing:
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
 
             def create_custom_forward(module):
                 def custom_forward(*inputs):

src/diffusers/models/autoencoders/autoencoder_kl_mochi.py

@@ -206,7 +206,7 @@ def forward(
         for i, (resnet, norm, attn) in enumerate(zip(self.resnets, self.norms, self.attentions)):
             conv_cache_key = f"resnet_{i}"
 
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module):
                     def create_forward(*inputs):
@@ -311,7 +311,7 @@ def forward(
         for i, (resnet, norm, attn) in enumerate(zip(self.resnets, self.norms, self.attentions)):
             conv_cache_key = f"resnet_{i}"
 
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module):
                     def create_forward(*inputs):
@@ -392,7 +392,7 @@ def forward(
         for i, resnet in enumerate(self.resnets):
             conv_cache_key = f"resnet_{i}"
 
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module):
                     def create_forward(*inputs):
@@ -529,7 +529,7 @@ def forward(
         hidden_states = self.proj_in(hidden_states)
         hidden_states = hidden_states.permute(0, 4, 1, 2, 3)
 
-        if self.training and self.gradient_checkpointing:
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
 
             def create_custom_forward(module):
                 def create_forward(*inputs):
@@ -646,7 +646,7 @@ def forward(
         hidden_states = self.conv_in(hidden_states)
 
         # 1. Mid
-        if self.training and self.gradient_checkpointing:
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
 
             def create_custom_forward(module):
                 def create_forward(*inputs):

src/diffusers/models/autoencoders/autoencoder_kl_temporal_decoder.py

@@ -95,7 +95,7 @@ def forward(
         sample = self.conv_in(sample)
 
         upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
-        if self.training and self.gradient_checkpointing:
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
 
             def create_custom_forward(module):
                 def custom_forward(*inputs):

src/diffusers/models/autoencoders/vae.py

@@ -142,7 +142,7 @@ def forward(self, sample: torch.Tensor) -> torch.Tensor:
 
         sample = self.conv_in(sample)
 
-        if self.training and self.gradient_checkpointing:
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
 
             def create_custom_forward(module):
                 def custom_forward(*inputs):
@@ -291,7 +291,7 @@ def forward(
         sample = self.conv_in(sample)
 
         upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
-        if self.training and self.gradient_checkpointing:
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
 
             def create_custom_forward(module):
                 def custom_forward(*inputs):
@@ -544,7 +544,7 @@ def forward(
         sample = self.conv_in(sample)
 
         upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
-        if self.training and self.gradient_checkpointing:
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
 
             def create_custom_forward(module):
                 def custom_forward(*inputs):
@@ -876,7 +876,7 @@ def __init__(
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         r"""The forward method of the `EncoderTiny` class."""
-        if self.training and self.gradient_checkpointing:
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
 
             def create_custom_forward(module):
                 def custom_forward(*inputs):
@@ -962,7 +962,7 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
         # Clamp.
         x = torch.tanh(x / 3) * 3
 
-        if self.training and self.gradient_checkpointing:
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
 
             def create_custom_forward(module):
                 def custom_forward(*inputs):

src/diffusers/models/controlnets/controlnet_flux.py

@@ -329,7 +329,7 @@ def forward(
 
         block_samples = ()
         for index_block, block in enumerate(self.transformer_blocks):
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module, return_dict=None):
                     def custom_forward(*inputs):
@@ -363,7 +363,7 @@ def custom_forward(*inputs):
 
         single_block_samples = ()
         for index_block, block in enumerate(self.single_transformer_blocks):
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module, return_dict=None):
                     def custom_forward(*inputs):

src/diffusers/models/controlnets/controlnet_sd3.py

@@ -324,7 +324,7 @@ def forward(
         block_res_samples = ()
 
         for block in self.transformer_blocks:
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module, return_dict=None):
                     def custom_forward(*inputs):

src/diffusers/models/controlnets/controlnet_xs.py

@@ -1466,7 +1466,7 @@ def custom_forward(*inputs):
                 h_ctrl = torch.cat([h_ctrl, b2c(h_base)], dim=1)
 
             # apply base subblock
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
                 ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                 h_base = torch.utils.checkpoint.checkpoint(
                     create_custom_forward(b_res),
@@ -1489,7 +1489,7 @@ def custom_forward(*inputs):
 
             # apply ctrl subblock
             if apply_control:
-                if self.training and self.gradient_checkpointing:
+                if torch.is_grad_enabled() and self.gradient_checkpointing:
                     ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                     h_ctrl = torch.utils.checkpoint.checkpoint(
                         create_custom_forward(c_res),
@@ -1898,7 +1898,7 @@ def maybe_apply_freeu_to_subblock(hidden_states, res_h_base):
             hidden_states, res_h_base = maybe_apply_freeu_to_subblock(hidden_states, res_h_base)
             hidden_states = torch.cat([hidden_states, res_h_base], dim=1)
 
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
                 ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                 hidden_states = torch.utils.checkpoint.checkpoint(
                     create_custom_forward(resnet),