Add option to use the fused optim version of ADamW

src/lerobot/datasets/dataset_tools.py

@@ -234,6 +234,8 @@ def merge_datasets(
     datasets: list[LeRobotDataset],
     output_repo_id: str,
     output_dir: str | Path | None = None,
+    data_files_size_in_mb: float | None = None,
+    video_files_size_in_mb: float | None = None,
 ) -> LeRobotDataset:
     """Merge multiple LeRobotDatasets into a single dataset.
 
@@ -257,6 +259,8 @@ def merge_datasets(
         aggr_repo_id=output_repo_id,
         roots=roots,
         aggr_root=output_dir,
+        data_files_size_in_mb=data_files_size_in_mb,
+        video_files_size_in_mb=video_files_size_in_mb,
     )
 
     merged_dataset = LeRobotDataset(
@@ -747,11 +751,11 @@ def _copy_and_reindex_videos(
                         f"videos/{video_key}/to_timestamp"
                     ]
             else:
-                # Build list of time ranges to keep, in sorted order.
+                # Episodes are already in order by old episode index (from episode_mapping iteration),
-                sorted_keep_episodes = sorted(episodes_in_file, key=lambda x: episode_mapping[x])
+                # which equals from_timestamp order since episodes are created sequentially.
                 episodes_to_keep_ranges: list[tuple[float, float]] = []
 
-                for old_idx in sorted_keep_episodes:
+                for old_idx in episodes_in_file:
                     src_ep = src_dataset.meta.episodes[old_idx]
                     from_ts = src_ep[f"videos/{video_key}/from_timestamp"]
                     to_ts = src_ep[f"videos/{video_key}/to_timestamp"]
@@ -781,7 +785,7 @@ def _copy_and_reindex_videos(
                 )
 
                 cumulative_ts = 0.0
-                for old_idx in sorted_keep_episodes:
+                for old_idx in episodes_in_file:
                     new_idx = episode_mapping[old_idx]
                     src_ep = src_dataset.meta.episodes[old_idx]
                     ep_length = src_ep["length"]

src/lerobot/datasets/lerobot_dataset.py

@@ -557,7 +557,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         episodes: list[int] | None = None,
         image_transforms: Callable | None = None,
         delta_timestamps: dict[str, list[float]] | None = None,
-        tolerance_s: float = 1e-4,
+        tolerance_s: float = 1e-2,
         revision: str | None = None,
         force_cache_sync: bool = False,
         download_videos: bool = True,

src/lerobot/optim/optimizers.py

@@ -81,10 +81,14 @@ class AdamWConfig(OptimizerConfig):
     eps: float = 1e-8
     weight_decay: float = 1e-2
     grad_clip_norm: float = 10.0
+    fused: bool = False  
 
     def build(self, params: dict) -> torch.optim.Optimizer:
         kwargs = asdict(self)
         kwargs.pop("grad_clip_norm")
+        # Fused optimizer only works on CUDA
+        if kwargs.get("fused") and not torch.cuda.is_available():
+            kwargs["fused"] = False
         return torch.optim.AdamW(params, **kwargs)
 
 

src/lerobot/policies/act/configuration_act.py

@@ -136,6 +136,7 @@ class ACTConfig(PreTrainedConfig):
     optimizer_lr: float = 1e-5
     optimizer_weight_decay: float = 1e-4
     optimizer_lr_backbone: float = 1e-5
+    optimizer_fused: bool = False  # Use CUDA fused AdamW kernel
 
     def __post_init__(self):
         super().__post_init__()
@@ -164,6 +165,7 @@ class ACTConfig(PreTrainedConfig):
         return AdamWConfig(
             lr=self.optimizer_lr,
             weight_decay=self.optimizer_weight_decay,
+            fused=self.optimizer_fused,
         )
 
     def get_scheduler_preset(self) -> None:

src/lerobot/policies/groot/configuration_groot.py

@@ -94,6 +94,7 @@ class GrootConfig(PreTrainedConfig):
     optimizer_betas: tuple[float, float] = (0.95, 0.999)
     optimizer_eps: float = 1e-8
     optimizer_weight_decay: float = 1e-5
+    optimizer_fused: bool = False  # Use CUDA fused AdamW kernel
     warmup_ratio: float = 0.05
     use_bf16: bool = True
 
@@ -174,6 +175,7 @@ class GrootConfig(PreTrainedConfig):
             betas=self.optimizer_betas,
             eps=self.optimizer_eps,
             weight_decay=self.optimizer_weight_decay,
+            fused=self.optimizer_fused,
         )
 
     def get_scheduler_preset(self) -> CosineDecayWithWarmupSchedulerConfig:

src/lerobot/policies/pi0/configuration_pi0.py

@@ -74,6 +74,7 @@ class PI0Config(PreTrainedConfig):
     gradient_checkpointing: bool = False  # Enable gradient checkpointing for memory optimization
     compile_model: bool = False  # Whether to use torch.compile for model optimization
     compile_mode: str = "max-autotune"  # Torch compile mode
+    optimizer_fused: bool = False  # Use CUDA fused AdamW kernel
     device: str | None = None  # Device to use for the model (None = auto-detect)
 
     # Optimizer settings: see openpi `AdamW``
@@ -141,6 +142,7 @@ class PI0Config(PreTrainedConfig):
             eps=self.optimizer_eps,
             weight_decay=self.optimizer_weight_decay,
             grad_clip_norm=self.optimizer_grad_clip_norm,
+            fused=self.optimizer_fused,
         )
 
     def get_scheduler_preset(self):

src/lerobot/policies/pi05/configuration_pi05.py

@@ -74,6 +74,7 @@ class PI05Config(PreTrainedConfig):
     gradient_checkpointing: bool = False  # Enable gradient checkpointing for memory optimization
     compile_model: bool = False  # Whether to use torch.compile for model optimization
     compile_mode: str = "max-autotune"  # Torch compile mode
+    optimizer_fused: bool = False  # Use CUDA fused AdamW kernel
     device: str | None = None  # Device to use for the model (None = auto-detect)
 
     # Optimizer settings: see openpi `AdamW`
@@ -141,6 +142,7 @@ class PI05Config(PreTrainedConfig):
             eps=self.optimizer_eps,
             weight_decay=self.optimizer_weight_decay,
             grad_clip_norm=self.optimizer_grad_clip_norm,
+            fused=self.optimizer_fused,
         )
 
     def get_scheduler_preset(self):

src/lerobot/policies/smolvla/configuration_smolvla.py

@@ -79,6 +79,7 @@ class SmolVLAConfig(PreTrainedConfig):
     optimizer_eps: float = 1e-8
     optimizer_weight_decay: float = 1e-10
     optimizer_grad_clip_norm: float = 10
+    optimizer_fused: bool = False 
 
     scheduler_warmup_steps: int = 1_000
     scheduler_decay_steps: int = 30_000
@@ -136,6 +137,7 @@ class SmolVLAConfig(PreTrainedConfig):
             eps=self.optimizer_eps,
             weight_decay=self.optimizer_weight_decay,
             grad_clip_norm=self.optimizer_grad_clip_norm,
+            fused=self.optimizer_fused,
         )
 
     def get_scheduler_preset(self):