feat(diffusion): add gradient checkpointing for memory optimization (#4127)

* feat(diffusion): add gradient checkpointing for memory optimization

Add gradient_checkpointing config option to DiffusionPolicy. When
enabled, wraps the UNet encoder, mid, and decoder residual blocks
with torch.utils.checkpoint.checkpoint to trade compute for memory.

Allows training with larger batch sizes or higher-resolution inputs
on memory-constrained GPUs. Disabled by default.

Usage: --policy.gradient_checkpointing=true

Part of the 0.6.0 roadmap item 3.3 (gradient checkpointing for all
policies).

* test(diffusion): verify gradient checkpointing parity

---------

Co-authored-by: Jash Shah <jashshah.999@gmail.com>

src/lerobot/policies/diffusion/configuration_diffusion.py

@@ -79,6 +79,8 @@ class DiffusionConfig(PreTrainedConfig):
         use_film_scale_modulation: FiLM (https://huggingface.co/papers/1709.07871) is used for the Unet conditioning.
             Bias modulation is used be default, while this parameter indicates whether to also use scale
             modulation.
+        gradient_checkpointing: Whether to checkpoint the Unet residual blocks during training. This reduces
+            activation memory at the cost of recomputing those blocks during the backward pass.
         noise_scheduler_type: Name of the noise scheduler to use. Supported options: ["DDPM", "DDIM"].
         num_train_timesteps: Number of diffusion steps for the forward diffusion schedule.
         beta_schedule: Name of the diffusion beta schedule as per DDPMScheduler from Hugging Face diffusers.
@@ -132,6 +134,7 @@ class DiffusionConfig(PreTrainedConfig):
     n_groups: int = 8
     diffusion_step_embed_dim: int = 128
     use_film_scale_modulation: bool = True
+    gradient_checkpointing: bool = False
     # Noise scheduler.
     noise_scheduler_type: str = "DDPM"
     num_train_timesteps: int = 100

src/lerobot/policies/diffusion/modeling_diffusion.py

@@ -31,6 +31,7 @@ import torch
 import torch.nn.functional as F  # noqa: N812
 import torchvision
 from torch import Tensor, nn
+from torch.utils.checkpoint import checkpoint
 
 from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_IMAGES, OBS_STATE
 from lerobot.utils.import_utils import _diffusers_available, require_package
@@ -727,22 +728,35 @@ class DiffusionConditionalUnet1d(nn.Module):
         else:
             global_feature = timesteps_embed
 
+        use_gc = self.config.gradient_checkpointing and self.training
+
         # Run encoder, keeping track of skip features to pass to the decoder.
         encoder_skip_features: list[Tensor] = []
         for resnet, resnet2, downsample in self.down_modules:
-            x = resnet(x, global_feature)
-            x = resnet2(x, global_feature)
+            if use_gc:
+                x = checkpoint(resnet, x, global_feature, use_reentrant=False)
+                x = checkpoint(resnet2, x, global_feature, use_reentrant=False)
+            else:
+                x = resnet(x, global_feature)
+                x = resnet2(x, global_feature)
             encoder_skip_features.append(x)
             x = downsample(x)
 
         for mid_module in self.mid_modules:
-            x = mid_module(x, global_feature)
+            if use_gc:
+                x = checkpoint(mid_module, x, global_feature, use_reentrant=False)
+            else:
+                x = mid_module(x, global_feature)
 
         # Run decoder, using the skip features from the encoder.
         for resnet, resnet2, upsample in self.up_modules:
             x = torch.cat((x, encoder_skip_features.pop()), dim=1)
-            x = resnet(x, global_feature)
-            x = resnet2(x, global_feature)
+            if use_gc:
+                x = checkpoint(resnet, x, global_feature, use_reentrant=False)
+                x = checkpoint(resnet2, x, global_feature, use_reentrant=False)
+            else:
+                x = resnet(x, global_feature)
+                x = resnet2(x, global_feature)
             x = upsample(x)
 
         x = self.final_conv(x)