Merge remote-tracking branch 'origin/main' into feat/smolvla-on-steerable

# Conflicts:
#	docs/source/annotation_pipeline.mdx
#	examples/annotations/run_hf_job.py
#	pyproject.toml
#	src/lerobot/annotations/steerable_pipeline/config.py
#	src/lerobot/annotations/steerable_pipeline/frames.py
#	src/lerobot/annotations/steerable_pipeline/modules/plan_subtasks_memory.py
#	src/lerobot/annotations/steerable_pipeline/vlm_client.py
#	src/lerobot/annotations/steerable_pipeline/writer.py
#	src/lerobot/datasets/__init__.py
#	src/lerobot/datasets/sampler.py
#	src/lerobot/scripts/lerobot_annotate.py
#	src/lerobot/scripts/lerobot_train.py
#	tests/annotations/test_frames.py
#	tests/annotations/test_modules.py
#	tests/annotations/test_writer.py
#	tests/datasets/test_sampler.py
#	tests/scripts/test_lerobot_annotate.py
#	uv.lock

tests/datasets/test_sampler.py

@@ -25,7 +25,7 @@ from datasets import Dataset  # noqa: E402
 from lerobot.datasets.io_utils import (
     hf_transform_to_torch,
 )
-from lerobot.datasets.sampler import EpisodeAwareSampler, WeightedEpisodeAwareSampler
+from lerobot.datasets.sampler import EpisodeAwareSampler, WeightedEpisodeAwareSampler, compute_sampler_state
 
 
 def calculate_episode_data_index(hf_dataset: Dataset) -> dict[str, torch.Tensor]:
@@ -114,6 +114,19 @@ def test_shuffle():
     assert set(sampler) == {0, 1, 2, 3, 4, 5}
 
 
+def test_shuffle_is_reproducible_across_instances():
+    # The order is a pure function of (seed, epoch), so two fresh samplers (e.g. two ranks)
+    # produce the same permutation without any generator synchronization.
+    sampler_a = EpisodeAwareSampler([0], [6], shuffle=True, seed=42)
+    sampler_b = EpisodeAwareSampler([0], [6], shuffle=True, seed=42)
+    epoch_0 = list(sampler_a)
+    assert list(sampler_b) == epoch_0
+    # Desyncing the global RNG must not affect the permutation.
+    sampler_c = EpisodeAwareSampler([0], [6], shuffle=True, seed=42)
+    torch.randperm(1000)  # consume global RNG, as rank-asymmetric code (e.g. eval) would
+    assert list(sampler_c) == epoch_0
+
+
 def test_negative_drop_first_frames_raises():
     with pytest.raises(ValueError, match="drop_n_first_frames must be >= 0"):
         EpisodeAwareSampler([0], [10], drop_n_first_frames=-1)
@@ -183,3 +196,85 @@ def test_weighted_sampler_zero_weights_fall_back_to_uniform():
 def test_weighted_sampler_rejects_short_weight_vector():
     with pytest.raises(ValueError, match="frame_weights"):
         WeightedEpisodeAwareSampler([0], [10], frame_weights=torch.ones(5))
+
+
+# --- seeded (seed, epoch) shuffling, resume, and state ---
+
+EPISODE_BOUNDS = ([0, 2, 3], [2, 3, 6])  # episodes of 2, 1 and 3 frames
+
+
+@pytest.mark.parametrize("num_frames", [1, 2, 3, 37, 64, 100])
+def test_deterministic_sampler_shuffle_is_permutation(num_frames):
+    for seed in (0, 1, 1234):
+        sampler = EpisodeAwareSampler([0], [num_frames], shuffle=True, seed=seed)
+        assert sorted(sampler) == list(range(num_frames))
+
+
+def test_deterministic_sampler_epochs_reproduce_and_differ():
+    sampler_a = EpisodeAwareSampler([0], [100], shuffle=True, seed=42)
+    sampler_b = EpisodeAwareSampler([0], [100], shuffle=True, seed=42)
+    epoch_0 = list(sampler_a)
+    assert list(sampler_b) == epoch_0  # same (seed, epoch) -> same order on any process
+    epoch_1 = list(sampler_a)  # __iter__ auto-advances the epoch
+    assert epoch_1 != epoch_0
+    assert sorted(epoch_1) == sorted(epoch_0)
+    sampler_a.set_epoch(0)
+    assert list(sampler_a) == epoch_0
+    assert list(EpisodeAwareSampler([0], [100], shuffle=True, seed=7)) != epoch_0
+
+
+def test_deterministic_sampler_resume_mid_epoch():
+    reference = EpisodeAwareSampler(*EPISODE_BOUNDS, shuffle=True, seed=42)
+    epoch_0 = list(reference)
+    epoch_1 = list(reference)
+    for start in (0, 1, 4, len(epoch_0)):
+        resumed = EpisodeAwareSampler(*EPISODE_BOUNDS, shuffle=True, seed=42)
+        resumed.load_state_dict({"epoch": 0, "start_index": start})
+        assert list(resumed) == epoch_0[start:]
+        # the resumed sampler continues into the same epoch 1 as the uninterrupted one
+        assert list(resumed) == epoch_1
+
+
+def test_deterministic_sampler_construction_stores_only_boundaries():
+    # Construction is O(num_episodes), not O(num_frames): a million-frame single episode
+    # instantiates from just its boundaries without materializing a per-frame index list.
+    num_frames = 1_000_000
+    sampler = EpisodeAwareSampler([0], [num_frames], shuffle=True, seed=0)
+    assert len(sampler) == num_frames
+    assert sampler._starts.shape == (1,) and sampler._cum_lengths.shape == (1,)
+
+
+def test_deterministic_sampler_resume_is_exact_at_scale():
+    # Seeded randperm makes resume sample-exact at non-trivial sizes: regenerating the epoch's
+    # permutation and slicing from the saved offset reproduces the remaining order exactly.
+    num_frames = 100_000
+    reference = EpisodeAwareSampler([0], [num_frames], shuffle=True, seed=0)
+    epoch_0 = list(reference)
+    assert sorted(epoch_0) == list(range(num_frames))
+    start = num_frames - 5
+    resumed = EpisodeAwareSampler([0], [num_frames], shuffle=True, seed=0)
+    resumed.load_state_dict({"epoch": 0, "start_index": start})
+    assert list(resumed) == epoch_0[start:]
+
+
+def test_compute_sampler_state():
+    # 100 frames, batch 10, 2 ranks -> 10 underlying batches, 5 per rank per epoch.
+    assert compute_sampler_state(step=0, num_frames=100, batch_size=10, num_processes=2) == {
+        "epoch": 0,
+        "start_index": 0,
+    }
+    # step 7 -> epoch 1, 2 per-rank batches in = 2 * 10 * 2 = 40 samples in
+    assert compute_sampler_state(step=7, num_frames=100, batch_size=10, num_processes=2) == {
+        "epoch": 1,
+        "start_index": 40,
+    }
+    # uneven epoch: 95 frames -> 10 underlying batches (last short), still 5 per rank
+    assert compute_sampler_state(step=12, num_frames=95, batch_size=10, num_processes=2) == {
+        "epoch": 2,
+        "start_index": 40,
+    }
+    # uneven sharding: 105 frames -> 11 underlying batches, 6 per rank (even_batches pads)
+    assert compute_sampler_state(step=11, num_frames=105, batch_size=10, num_processes=2) == {
+        "epoch": 1,
+        "start_index": 100,
+    }