from collections import Counter import torch from psi.data.sampler import TaskMixtureSampler from psi.data.seed.omni_oss_sampler import OssShardBatch class _FakeShardAwareDataset: def __init__(self): self.calls = 0 def sample_shard_aware_indices(self, batch_size, generator): self.calls += 1 return list(range(10, 10 + batch_size)) def _dataset_sequence(sampler): sequence = [] for batch in sampler: dataset_ids = {item[0] for item in batch} assert len(dataset_ids) == 1 sequence.append(next(iter(dataset_ids))) return sequence def test_task_mixture_sampler_batches_are_homogeneous(monkeypatch): monkeypatch.setenv("RANK", "0") monkeypatch.setenv("WORLD_SIZE", "1") monkeypatch.setenv("PSI_TASK_MIXTURE_SHARD_AWARE", "0") sampler = TaskMixtureSampler( dataset_lens=[100, 50, 30], mixture_ratios=[3, 1, 0], num_samples_per_epoch=64, batch_size=4, seed=7, ) batches = list(sampler) assert len(batches) == 16 for batch in batches: assert len(batch) == 4 assert len({dataset_id for dataset_id, _ in batch}) == 1 def test_task_mixture_sampler_respects_batch_ratios(monkeypatch): monkeypatch.setenv("RANK", "0") monkeypatch.setenv("WORLD_SIZE", "1") monkeypatch.setenv("PSI_TASK_MIXTURE_SHARD_AWARE", "0") sampler = TaskMixtureSampler( dataset_lens=[100, 50], mixture_ratios=[3, 1], num_samples_per_epoch=64, batch_size=4, seed=11, ) counts = Counter(_dataset_sequence(sampler)) assert counts == Counter({0: 12, 1: 4}) def test_task_mixture_sampler_distributed_schedule_matches_global(monkeypatch): monkeypatch.setenv("PSI_TASK_MIXTURE_SYNC_RANKS", "0") common_kwargs = dict( dataset_lens=[100, 50, 30], mixture_ratios=[5, 3, 2], num_samples_per_epoch=64, batch_size=4, seed=19, ) monkeypatch.setenv("RANK", "0") monkeypatch.setenv("WORLD_SIZE", "1") global_sequence = _dataset_sequence(TaskMixtureSampler(**common_kwargs)) monkeypatch.setenv("RANK", "0") monkeypatch.setenv("WORLD_SIZE", "2") rank0_sequence = _dataset_sequence(TaskMixtureSampler(**common_kwargs)) monkeypatch.setenv("RANK", "1") monkeypatch.setenv("WORLD_SIZE", "2") rank1_sequence = _dataset_sequence(TaskMixtureSampler(**common_kwargs)) merged_sequence = [] for batch_idx in range(len(global_sequence)): local_idx = batch_idx // 2 merged_sequence.append(rank0_sequence[local_idx] if batch_idx % 2 == 0 else rank1_sequence[local_idx]) assert merged_sequence == global_sequence def test_task_mixture_sampler_can_sync_tasks_across_ranks(monkeypatch): common_kwargs = dict( dataset_lens=[100, 50, 30], mixture_ratios=[5, 3, 2], num_samples_per_epoch=64, batch_size=4, seed=23, ) monkeypatch.setenv("PSI_TASK_MIXTURE_SYNC_RANKS", "1") monkeypatch.setenv("WORLD_SIZE", "2") monkeypatch.setenv("RANK", "0") rank0_batches = list(TaskMixtureSampler(**common_kwargs)) rank0_sequence = _dataset_sequence(rank0_batches) monkeypatch.setenv("RANK", "1") rank1_batches = list(TaskMixtureSampler(**common_kwargs)) rank1_sequence = _dataset_sequence(rank1_batches) assert rank0_sequence == rank1_sequence assert all( {sample_id for _, sample_id in batch0} != {sample_id for _, sample_id in batch1} for batch0, batch1 in zip(rank0_batches, rank1_batches) ) def test_task_mixture_sampler_balances_equal_tasks_across_ranks(monkeypatch): common_kwargs = dict( dataset_lens=[100, 100, 100, 100], mixture_ratios=[1, 1, 1, 1], num_samples_per_epoch=64, batch_size=4, seed=23, ) monkeypatch.setenv("PSI_TASK_MIXTURE_BALANCED_RANKS", "1") monkeypatch.setenv("PSI_TASK_MIXTURE_SHARD_AWARE", "0") monkeypatch.setenv("WORLD_SIZE", "8") rank_sequences = [] for rank in range(8): monkeypatch.setenv("RANK", str(rank)) batches = list(TaskMixtureSampler(**common_kwargs)) assert batches task_ids = [{item[0] for item in batch} for batch in batches] assert all(len(ids) == 1 for ids in task_ids) rank_sequences.append([next(iter(ids)) for ids in task_ids]) # Every local step has exactly two ranks assigned to each enabled task. for local_step in range(len(rank_sequences[0])): assert Counter(sequence[local_step] for sequence in rank_sequences) == Counter( {0: 2, 1: 2, 2: 2, 3: 2} ) def test_task_mixture_sampler_assigns_exact_weighted_rank_counts(monkeypatch): common_kwargs = dict( dataset_lens=[100, 100, 100, 100], mixture_ratios=[3, 1, 3, 1], num_samples_per_epoch=64, batch_size=4, seed=29, ) monkeypatch.setenv("PSI_TASK_MIXTURE_BALANCED_RANKS", "1") monkeypatch.setenv("PSI_TASK_MIXTURE_SHARD_AWARE", "0") monkeypatch.setenv("WORLD_SIZE", "8") rank_sequences = [] for rank in range(8): monkeypatch.setenv("RANK", str(rank)) batches = list(TaskMixtureSampler(**common_kwargs)) assert batches rank_sequences.append(_dataset_sequence(batches)) for local_step in range(len(rank_sequences[0])): assert Counter(sequence[local_step] for sequence in rank_sequences) == Counter( {0: 3, 1: 1, 2: 3, 3: 1} ) def test_task_mixture_sampler_uses_only_residual_flex_ranks_for_fractional_weights(monkeypatch): common_kwargs = dict( dataset_lens=[19, 6, 19, 56], mixture_ratios=[19, 6, 19, 56], num_samples_per_epoch=8000, batch_size=10, seed=31, ) monkeypatch.setenv("PSI_TASK_MIXTURE_BALANCED_RANKS", "1") monkeypatch.setenv("PSI_TASK_MIXTURE_SHARD_AWARE", "0") monkeypatch.setenv("PSI_TASK_MIXTURE_BLOCK_BATCHES", "8") monkeypatch.setenv("WORLD_SIZE", "8") rank_sequences = [] fixed_ranks = 0 for rank in range(8): monkeypatch.setenv("RANK", str(rank)) sequence = _dataset_sequence(TaskMixtureSampler(**common_kwargs)) rank_sequences.append(sequence) fixed_ranks += int(len(set(sequence)) == 1) # floor([1.52, 0.48, 1.52, 4.48]) keeps six core ranks resident; # only two flex ranks realize the residual proportions. assert fixed_ranks == 6 assert Counter(task for sequence in rank_sequences for task in sequence) == Counter( {0: 152, 1: 48, 2: 152, 3: 448} ) def test_fractional_flex_rank_initializes_all_tasks_before_sampling(monkeypatch): class _InitAwareDataset: def __init__(self): self.batch_sizes = [] def sample_shard_aware_indices(self, batch_size, generator, lane_id=0): del generator, lane_id self.batch_sizes.append(batch_size) return list(range(batch_size)) monkeypatch.setenv("PSI_TASK_MIXTURE_BALANCED_RANKS", "1") monkeypatch.setenv("PSI_TASK_MIXTURE_SHARD_AWARE", "1") monkeypatch.setenv("WORLD_SIZE", "8") datasets = [_InitAwareDataset() for _ in range(4)] common_kwargs = dict( dataset_lens=[19, 6, 19, 56], datasets=datasets, mixture_ratios=[19, 6, 19, 56], num_samples_per_epoch=800, batch_size=10, seed=41, ) flex_rank = None for rank in range(8): monkeypatch.setenv("RANK", str(rank)) plan = TaskMixtureSampler(**common_kwargs)._balanced_rank_plan() if plan is not None and plan[1] is not None: flex_rank = rank break assert flex_rank is not None monkeypatch.setenv("RANK", str(flex_rank)) list(TaskMixtureSampler(**common_kwargs)) assert all(dataset.batch_sizes[0] == 0 for dataset in datasets) assert all(any(batch_size > 0 for batch_size in dataset.batch_sizes) for dataset in datasets) def test_task_mixture_sampler_uses_shard_aware_indices_only_when_enabled(monkeypatch): monkeypatch.setenv("RANK", "0") monkeypatch.setenv("WORLD_SIZE", "1") dataset = _FakeShardAwareDataset() common_kwargs = dict( dataset_lens=[100], datasets=[dataset], mixture_ratios=[1], num_samples_per_epoch=16, batch_size=4, seed=31, ) monkeypatch.setenv("PSI_TASK_MIXTURE_SHARD_AWARE", "0") default_batches = list(TaskMixtureSampler(**common_kwargs)) assert dataset.calls == 0 assert any(sample_id < 10 or sample_id >= 14 for batch in default_batches for _, sample_id in batch) monkeypatch.setenv("PSI_TASK_MIXTURE_SHARD_AWARE", "1") shard_batches = list(TaskMixtureSampler(**common_kwargs)) assert dataset.calls == len(shard_batches) assert all([sample_id for _, sample_id in batch] == [10, 11, 12, 13] for batch in shard_batches) def test_task_mixture_sampler_emits_oss_hints_and_worker_lanes(monkeypatch): class _FakeOssDataset: def __init__(self): self.lanes = [] def sample_shard_aware_indices(self, batch_size, generator, lane_id=0): self.lanes.append(lane_id) return OssShardBatch( indices=list(range(lane_id * 100, lane_id * 100 + batch_size)), current_shard=f"/oss/current-{lane_id}.safetensors", next_shard=f"/oss/next-{lane_id}.safetensors", shard_id=lane_id, ) monkeypatch.setenv("RANK", "0") monkeypatch.setenv("WORLD_SIZE", "1") monkeypatch.setenv("PSI_TASK_MIXTURE_SHARD_AWARE", "1") monkeypatch.setenv("PSI_TASK_MIXTURE_BLOCK_BATCHES", "4") monkeypatch.setenv("PSI_TASK_MIXTURE_RANK_LOCAL_BLOCKS", "1") monkeypatch.setenv("PSI_OMNI_OSS_WORKER_LANES", "2") dataset = _FakeOssDataset() sampler = TaskMixtureSampler( dataset_lens=[1000], datasets=[dataset], mixture_ratios=[1], num_samples_per_epoch=32, batch_size=4, seed=37, ) batches = list(sampler) assert dataset.lanes == [0, 1, 0, 1, 0, 1, 0, 1] assert all(len(item) == 4 for batch in batches for item in batch) assert all(item[2] == f"/oss/current-{batch_id % 2}.safetensors" for batch_id, batch in enumerate(batches) for item in batch) def test_task_mixture_sampler_fast_forward_matches_exact_continuation(monkeypatch): class _StatefulShardDataset: def __init__(self, dataset_id): self.dataset_id = dataset_id self.lane_calls = Counter() self.getitem_calls = 0 def sample_shard_aware_indices(self, batch_size, generator, lane_id=0): call = self.lane_calls[lane_id] self.lane_calls[lane_id] += 1 random_part = torch.randint(0, 1000, (batch_size,), generator=generator).tolist() return [ self.dataset_id * 100_000 + lane_id * 10_000 + call * 1000 + value for value in random_part ] def __getitem__(self, index): self.getitem_calls += 1 return index monkeypatch.setenv("WORLD_SIZE", "8") monkeypatch.setenv("PSI_TASK_MIXTURE_BALANCED_RANKS", "1") monkeypatch.setenv("PSI_TASK_MIXTURE_SHARD_AWARE", "1") monkeypatch.setenv("PSI_TASK_MIXTURE_BLOCK_BATCHES", "8") monkeypatch.setenv("PSI_OMNI_OSS_WORKER_LANES", "3") common_kwargs = dict( dataset_lens=[190_000, 60_000, 190_000, 560_000], mixture_ratios=[19, 6, 19, 56], num_samples_per_epoch=8 * 40 * 4, batch_size=4, seed=43, ) offset = 17 # Cover both fixed and residual-flex ranks from the production schedule. for rank in range(8): monkeypatch.setenv("RANK", str(rank)) baseline_datasets = [_StatefulShardDataset(i) for i in range(4)] baseline = TaskMixtureSampler(datasets=baseline_datasets, **common_kwargs) baseline.set_epoch(1) expected = list(baseline)[offset:] resumed_datasets = [_StatefulShardDataset(i) for i in range(4)] resumed = TaskMixtureSampler(datasets=resumed_datasets, **common_kwargs) resumed.set_epoch(1) reports = [] assert resumed.fast_forward(offset, progress_callback=reports.append) == offset assert reports[-1] == offset assert list(resumed) == expected assert sum(dataset.getitem_calls for dataset in resumed_datasets) == 0