[Feature] Add _Contiguous module and reshape improvements to encoders/decoders

- Add _Contiguous helper module for torch.compile inductor compatibility
- Refactor ObsEncoder.forward and ObsDecoder.forward to use flatten/unflatten with contiguous()
- Add _maybe_record_function_decorator for profiling


ghstack-source-id: 86ef6a9
Pull-Request: #3306

Author: vmoens

torchrl/modules/models/model_based.py

@@ -18,13 +18,25 @@
 
 # from torchrl.modules.tensordict_module.rnn import GRUCell
 from torch.nn import GRUCell
-from torchrl._utils import timeit
+from torchrl._utils import _maybe_record_function_decorator
 
 from torchrl.modules.models.models import MLP
 
 UNSQUEEZE_RNN_INPUT = version.parse(torch.__version__) < version.parse("1.11")
 
 
+class _Contiguous(nn.Module):
+    """Helper module that makes a tensor contiguous.
+
+    This is useful inside nn.Sequential for torch.compile inductor compatibility.
+    Inductor sometimes needs explicit contiguous() calls after reshape operations
+    for efficient memory layout.
+    """
+
+    def forward(self, x):
+        return x.contiguous()
+
+
 class DreamerActor(nn.Module):
     """Dreamer actor network.
 
@@ -128,16 +140,16 @@ def __init__(self, channels=32, num_layers=4, in_channels=None, depth=None, devi
             k = k * 2
         self.encoder = nn.Sequential(*layers)
 
+    @_maybe_record_function_decorator("ObsEncoder.forward")
     def forward(self, observation):
         *batch_sizes, C, H, W = observation.shape
-        if len(batch_sizes) == 0:
-            end_dim = 0
-        else:
-            end_dim = len(batch_sizes) - 1
-        observation = torch.flatten(observation, start_dim=0, end_dim=end_dim)
+        # Flatten all batch dimensions into one for conv
+        # Use contiguous() for inductor compatibility
+        observation = observation.flatten(0, len(batch_sizes) - 1 if batch_sizes else 0).contiguous()
         obs_encoded = self.encoder(observation)
-        latent = obs_encoded.reshape(*batch_sizes, -1)
-        return latent
+        # Reshape back to original batch dims + latent
+        latent = obs_encoded.unflatten(0, batch_sizes) if batch_sizes else obs_encoded
+        return latent.reshape(*batch_sizes, -1).contiguous()
 
 
 class ObsDecoder(nn.Module):
@@ -213,14 +225,18 @@ def __init__(self, channels=32, num_layers=4, kernel_sizes=None, latent_dim=None
         self.decoder = nn.Sequential(*layers)
         self._depth = channels
 
+    @_maybe_record_function_decorator("ObsDecoder.forward")
     def forward(self, state, rnn_hidden):
+        # Concatenate and project to latent space
         latent = self.state_to_latent(torch.cat([state, rnn_hidden], dim=-1))
         *batch_sizes, D = latent.shape
-        latent = latent.view(-1, D, 1, 1)
+        # Flatten batch dimensions and reshape for conv
+        latent = latent.flatten(0, len(batch_sizes) - 1 if batch_sizes else 0).unsqueeze(-1).unsqueeze(-1).contiguous()
         obs_decoded = self.decoder(latent)
         _, C, H, W = obs_decoded.shape
-        obs_decoded = obs_decoded.view(*batch_sizes, C, H, W)
-        return obs_decoded
+        # Unflatten back to original batch dims
+        obs_decoded = obs_decoded.unflatten(0, batch_sizes) if batch_sizes else obs_decoded
+        return obs_decoded.contiguous()
 
 
 class RSSMRollout(TensorDictModuleBase):