Introduce Dataset Optimizer (#18788)

Co-authored-by: thomas <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
(cherry picked from commit 142977d)

Author: 3 people

requirements/app/app.txt

@@ -1,4 +1,4 @@
-lightning-cloud ==0.5.39  # Must be pinned to ensure compatibility
+lightning-cloud ==0.5.41  # Must be pinned to ensure compatibility
 packaging
 typing-extensions >=4.0.0, <4.8.0
 deepdiff >=5.7.0, <6.6.0

src/lightning/data/cache/__init__.py

@@ -13,5 +13,6 @@
 
 from lightning.data.cache.cache import Cache
 from lightning.data.cache.dataloader import LightningDataLoader
+from lightning.data.cache.dataset_optimizer import DatasetOptimizer
 
-__all__ = ["Cache", "LightningDataLoader"]
+__all__ = ["Cache", "DatasetOptimizer", "LightningDataLoader"]

src/lightning/data/cache/cache.py

@@ -15,7 +15,7 @@
 import os
 from typing import Any, Dict, List, Optional, Tuple, Union
 
-from lightning.data.cache.constants import _INDEX_FILENAME, _TORCH_2_1_0_AVAILABLE
+from lightning.data.cache.constants import _INDEX_FILENAME, _TORCH_GREATER_EQUAL_2_1_0
 from lightning.data.cache.reader import BinaryReader
 from lightning.data.cache.sampler import ChunkedIndex
 from lightning.data.cache.writer import BinaryWriter
@@ -46,11 +46,13 @@ def __init__(
 
         """
         super().__init__()
-        if not _TORCH_2_1_0_AVAILABLE:
+        if not _TORCH_GREATER_EQUAL_2_1_0:
             raise ModuleNotFoundError("PyTorch version 2.1 or higher is required to use the cache.")
-        self._writer = BinaryWriter(cache_dir, chunk_size=chunk_size, chunk_bytes=chunk_bytes, compression=compression)
-        self._reader = BinaryReader(cache_dir, remote_dir=remote_dir, compression=compression)
-        self._cache_dir = cache_dir
+        self._writer = BinaryWriter(
+            str(cache_dir), chunk_size=chunk_size, chunk_bytes=chunk_bytes, compression=compression
+        )
+        self._reader = BinaryReader(str(cache_dir), remote_dir=remote_dir, compression=compression)
+        self._cache_dir = str(cache_dir)
         self._is_done = False
         self._distributed_env = _DistributedEnv.detect()
 
@@ -66,19 +68,27 @@ def __setitem__(self, index: int, data: Any) -> None:
         """Store an item in the writer."""
         self._writer[index] = data
 
+    def _add_item(self, index: int, data: Any) -> Optional[str]:
+        """Store an item in the writer and optionally return the chunk path."""
+        return self._writer.add_item(index, data)
+
     def __getitem__(self, index: Union[int, ChunkedIndex]) -> Dict[str, Any]:
         """Read an item in the reader."""
         if isinstance(index, int):
             index = ChunkedIndex(index, self._get_chunk_index_from_index(index))
         return self._reader.read(index)
 
-    def done(self) -> None:
+    def done(self) -> Optional[List[str]]:
+        """Inform the writer the chunking phase is finished."""
+        return self._writer.done()
+
+    def merge(self, num_workers: int = 1, node_rank: Optional[int] = None) -> None:
         """Inform the writer the chunking phase is finished."""
-        self._writer.done()
+        self._writer.merge(num_workers, node_rank=node_rank)
 
-    def merge(self, num_workers: int = 1) -> None:
+    def _merge_no_wait(self, node_rank: Optional[int] = None) -> None:
         """Inform the writer the chunking phase is finished."""
-        self._writer.merge(num_workers)
+        self._writer._merge_no_wait(node_rank=node_rank)
 
     def __len__(self) -> int:
         return self._reader.get_length()

src/lightning/data/cache/config.py

@@ -15,11 +15,11 @@
 import os
 from typing import Any, Dict, List, Optional, Tuple
 
-from lightning.data.cache.constants import _INDEX_FILENAME, _TORCH_2_1_0_AVAILABLE
+from lightning.data.cache.constants import _INDEX_FILENAME, _TORCH_GREATER_EQUAL_2_1_0
 from lightning.data.cache.downloader import get_downloader_cls
 from lightning.data.cache.sampler import ChunkedIndex
 
-if _TORCH_2_1_0_AVAILABLE:
+if _TORCH_GREATER_EQUAL_2_1_0:
     from torch.utils._pytree import treespec_loads
 
 
@@ -54,6 +54,7 @@ def __init__(self, cache_dir: str, remote_dir: Optional[str]):
             if (end - start) != chunk["chunk_size"]:
                 raise Exception(
                     "The config intervals doesn't match the number of samples. This shouldn't have happened."
+                    f" Found {end} {start} {chunk['chunk_size']}"
                 )
             self._intervals.append((chunk["interval"][0], chunk["interval"][1]))
 

src/lightning/data/cache/constants.py

@@ -15,7 +15,10 @@
 
 _INDEX_FILENAME = "index.json"
 _DEFAULT_CHUNK_BYTES = 1 << 26  # 64M B
+_DEFAULT_FAST_DEV_RUN_ITEMS = 10
 
 # This is required for full pytree serialization / deserialization support
-_TORCH_2_1_0_AVAILABLE = RequirementCache("torch>=2.1.0")
+_TORCH_GREATER_EQUAL_2_1_0 = RequirementCache("torch>=2.1.0")
 _VIZ_TRACKER_AVAILABLE = RequirementCache("viztracer")
+_LIGHTNING_CLOUD_GREATER_EQUAL_0_5_41 = RequirementCache("lightning-cloud>=0.5.41")
+_BOTO3_AVAILABLE = RequirementCache("boto3")

src/lightning/data/cache/dataloader.py

@@ -32,11 +32,11 @@
 from torch.utils.data.sampler import BatchSampler, Sampler
 
 from lightning.data.cache import Cache
-from lightning.data.cache.constants import _DEFAULT_CHUNK_BYTES, _TORCH_2_1_0_AVAILABLE, _VIZ_TRACKER_AVAILABLE
+from lightning.data.cache.constants import _DEFAULT_CHUNK_BYTES, _TORCH_GREATER_EQUAL_2_1_0, _VIZ_TRACKER_AVAILABLE
 from lightning.data.cache.sampler import CacheBatchSampler
-from lightning.data.datasets.env import _DistributedEnv, _WorkerEnv
+from lightning.data.datasets.env import _DistributedEnv
 
-if _TORCH_2_1_0_AVAILABLE:
+if _TORCH_GREATER_EQUAL_2_1_0:
     from torch.utils._pytree import tree_flatten
 
 logger = logging.Logger(__name__)
@@ -154,13 +154,27 @@ def __init__(self, global_rank: int, profile: bool = False) -> None:
         self._global_rank = global_rank
         self._profile = profile
 
-    def __call__(self, dataset_kind: _DatasetKind, *args: Any, **kwargs: Any) -> None:
+    def __call__(
+        self,
+        dataset_kind: Any,
+        dataset: Any,
+        index_queue: Any,
+        data_queue: Any,
+        done_event: Any,
+        auto_collation: Any,
+        collate_fn: Any,
+        drop_last: Any,
+        base_seed: Any,
+        init_fn: Any,
+        worker_id: Any,
+        *args: Any,
+        **kwargs: Any,
+    ) -> None:
         from torch.utils.data._utils import worker
 
         from lightning.data.cache.cache import Cache
 
-        rank = _WorkerEnv.detect().rank
-        enable_profiling = self._global_rank == 0 and rank == 0 and _VIZ_TRACKER_AVAILABLE and self._profile
+        enable_profiling = self._global_rank == 0 and worker_id == 0 and _VIZ_TRACKER_AVAILABLE and self._profile
 
         if enable_profiling:
             from viztracer import VizTracer
@@ -180,7 +194,21 @@ def create_fetcher_fn(*args: Any, **kwargs: Any) -> "_BaseDatasetFetcher":
 
         _DatasetKind.create_fetcher = create_fetcher_fn  # type: ignore
 
-        reloaded_worker._worker_loop(dataset_kind, *args, **kwargs)
+        reloaded_worker._worker_loop(
+            dataset_kind,
+            dataset,
+            index_queue,
+            data_queue,
+            done_event,
+            auto_collation,
+            collate_fn,
+            drop_last,
+            base_seed,
+            init_fn,
+            worker_id,
+            *args,
+            **kwargs,
+        )
 
         if dataset_kind == _DatasetKind.Map:
             assert fetcher