fix(train): synchronize EpisodeAwareSampler shuffling across ranks and gate dataset download per node (#3768)

* fix(datasets): expose a generator on EpisodeAwareSampler for distributed shuffle sync

In distributed training, accelerate can only synchronize the shuffle
permutation across ranks when the sampler exposes a generator attribute.
EpisodeAwareSampler shuffled via the global torch RNG, so disjoint batch
shards relied on every rank's global CPU RNG staying in lockstep forever;
any rank-asymmetric RNG consumption (e.g. eval rollouts on the main
process only) silently desynced the permutations and ranks trained on
overlapping/missing samples.

* fix(train): seed sampler generator and gate dataset download per node

- Pass a generator seeded with cfg.seed to EpisodeAwareSampler so
  accelerator.prepare registers it as the synchronized RNG and the
  shuffle order is reproducible.
- Gate the initial make_dataset call on is_local_main_process instead of
  is_main_process: the global main process only exists on node 0, so on
  every other node all local ranks were downloading the dataset and
  building the Arrow cache concurrently.

src/lerobot/datasets/sampler.py

@@ -30,6 +30,7 @@ class EpisodeAwareSampler:
         drop_n_first_frames: int = 0,
         drop_n_last_frames: int = 0,
         shuffle: bool = False,
+        generator: torch.Generator | None = None,
     ):
         """Sampler that optionally incorporates episode boundary information.
 
@@ -41,6 +42,10 @@ class EpisodeAwareSampler:
             drop_n_first_frames: Number of frames to drop from the start of each episode.
             drop_n_last_frames: Number of frames to drop from the end of each episode.
             shuffle: Whether to shuffle the indices.
+            generator: Generator used for shuffling. Exposing this attribute (even when None) lets
+                       `accelerate` register it as the synchronized RNG in distributed training, so
+                       every rank draws the same permutation and batch shards stay disjoint. When
+                       None, shuffling falls back to the global torch RNG.
         """
         if drop_n_first_frames < 0:
             raise ValueError(f"drop_n_first_frames must be >= 0, got {drop_n_first_frames}")
@@ -73,10 +78,11 @@ class EpisodeAwareSampler:
 
         self.indices = indices
         self.shuffle = shuffle
+        self.generator = generator
 
     def __iter__(self) -> Iterator[int]:
         if self.shuffle:
-            for i in torch.randperm(len(self.indices)):
+            for i in torch.randperm(len(self.indices), generator=self.generator):
                 yield self.indices[i]
         else:
             for i in self.indices:

src/lerobot/scripts/lerobot_train.py

@@ -232,15 +232,18 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         torch.backends.cudnn.benchmark = True
     torch.backends.cuda.matmul.allow_tf32 = True
 
-    # Dataset loading synchronization: main process downloads first to avoid race conditions
-    if is_main_process:
-        logging.info("Creating dataset")
+    # Dataset loading synchronization: each node's local main process downloads first to avoid
+    # race conditions (the global main process only exists on node 0, so gating on it would let
+    # all ranks of the other nodes download and build the Arrow cache concurrently).
+    if accelerator.is_local_main_process:
+        if is_main_process:
+            logging.info("Creating dataset")
         dataset = make_dataset(cfg)
 
     accelerator.wait_for_everyone()
 
-    # Now all other processes can safely load the dataset
-    if not is_main_process:
+    # Now all other processes can safely load the dataset from the local cache
+    if not accelerator.is_local_main_process:
         dataset = make_dataset(cfg)
 
     # Create environment used for evaluating checkpoints during training on simulation data.
@@ -386,12 +389,19 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
     # create dataloader for offline training
     if hasattr(active_cfg, "drop_n_last_frames"):
         shuffle = False
+        # A dedicated generator (rather than the global torch RNG) lets accelerator.prepare
+        # synchronize the shuffle permutation across ranks, keeping batch shards disjoint even
+        # when ranks consume the global RNG asymmetrically (e.g. eval on the main process only).
+        sampler_generator = torch.Generator()
+        if cfg.seed is not None:
+            sampler_generator.manual_seed(cfg.seed)
         sampler = EpisodeAwareSampler(
             dataset.meta.episodes["dataset_from_index"],
             dataset.meta.episodes["dataset_to_index"],
             episode_indices_to_use=dataset.episodes,
             drop_n_last_frames=active_cfg.drop_n_last_frames,
             shuffle=True,
+            generator=sampler_generator,
         )
     else:
         shuffle = True