Fix oversized items not fitting into a chunk (#18938)

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
(cherry picked from commit 0e7a3b0)

Author: 2 people

src/lightning/data/streaming/writer.py

@@ -13,6 +13,7 @@
 
 import json
 import os
+import warnings
 from dataclasses import dataclass
 from time import sleep
 from typing import Any, Dict, List, Optional, Tuple, Union
@@ -24,6 +25,7 @@
 from lightning.data.streaming.compression import _COMPRESSORS, Compressor
 from lightning.data.streaming.constants import _INDEX_FILENAME, _TORCH_GREATER_EQUAL_2_1_0
 from lightning.data.streaming.serializers import _SERIALIZERS, Serializer
+from lightning.data.utilities.format import _human_readable_bytes
 
 if _TORCH_GREATER_EQUAL_2_1_0:
     from torch.utils._pytree import PyTree, tree_flatten, treespec_dumps
@@ -42,20 +44,16 @@
 
 
 def _convert_bytes_to_int(bytes_str: str) -> int:
-    """Convert human readable byte format to an integer."""
+    """Convert human-readable byte format to an integer."""
     for suffix in _FORMAT_TO_RATIO:
         bytes_str = bytes_str.lower().strip()
         if bytes_str.lower().endswith(suffix):
             try:
                 return int(float(bytes_str[0 : -len(suffix)]) * _FORMAT_TO_RATIO[suffix])
             except ValueError:
                 raise ValueError(
-                    "".join(
-                        [
-                            f"Unsupported value/suffix {bytes_str}. Supported suffix are ",
-                            f'{["b"] + list(_FORMAT_TO_RATIO.keys())}.',
-                        ]
-                    )
+                    f"Unsupported value/suffix {bytes_str}. Supported suffix are "
+                    f'{["b"] + list(_FORMAT_TO_RATIO.keys())}.'
                 )
     raise ValueError(f"The supported units are {_FORMAT_TO_RATIO.keys()}")
 
@@ -212,39 +210,42 @@ def _serialize(self, item: Any, sizes: List[int], data: List[bytes]) -> str:
 
     def _create_chunk(self, filename: str, on_done: bool = False) -> bytes:
         """Create a binary chunk from all the binarized items."""
+        items = []
+
         if on_done:
             indices = sorted(self._serialized_items.keys())
             for i in range(len(indices) - 1):
                 assert indices[i] == indices[i + 1] - 1, indices
-            min_index = indices[0]
-            max_index = indices[-1] + 1
-            num_items = np.uint32(max_index - min_index)
             items = [self._serialized_items.pop(index) for index in indices]
         else:
             assert self._max_index is not None, (self._max_index, self._min_index)
             assert self._min_index is not None, (self._max_index, self._min_index)
-            num_items = np.uint32(self._max_index - self._min_index)
-            items = [self._serialized_items.pop(index) for index in range(self._min_index, self._max_index)]
-            min_index = self._min_index
-            max_index = self._max_index
+            if self._max_index == self._min_index:
+                # A single item is larger than the target chunk size; allow the chunk to be bigger than the target size
+                items.append(self._serialized_items.pop(self._max_index))
+            items.extend(self._serialized_items.pop(index) for index in range(self._min_index, self._max_index))
 
         if len(items) == 0:
             raise RuntimeError(
                 "The items shouldn't have an empty length. Something went wrong."
                 f" Found {self._pretty_serialized_items()} with boundaries: {self._min_index}, {self._max_index}."
             )
 
+        num_items = np.uint32(len(items))
         sizes = list(map(len, items))
         offsets = np.array([0] + sizes).cumsum().astype(np.uint32)
         offsets += len(num_items.tobytes()) + len(offsets.tobytes())
         sample_data = b"".join([item.data for item in items])
         data = num_items.tobytes() + offsets.tobytes() + sample_data
-        offsets = offsets.tolist()
 
         current_chunk_bytes = sum([item.bytes for item in items])
 
-        if self._chunk_bytes:
-            assert current_chunk_bytes <= self._chunk_bytes
+        if self._chunk_bytes and current_chunk_bytes > self._chunk_bytes:
+            warnings.warn(
+                f"An item was larger than the target chunk size ({_human_readable_bytes(self._chunk_bytes)})."
+                f" The current chunk will be {_human_readable_bytes(current_chunk_bytes)} in size.",
+                UserWarning,
+            )
 
         if self._chunk_size:
             assert num_items.item() <= self._chunk_size
@@ -308,6 +309,7 @@ def add_item(self, index: int, items: Any) -> Optional[str]:
             return filepath
 
     def _should_write(self) -> bool:
+        # TODO: Misleading method name, it modifies `self._min_index` and `self._max_index`!
         if not self._serialized_items:
             return False
         indexes = list(self._serialized_items.keys())

src/lightning/data/utilities/__init__.py

@@ -0,0 +1,6 @@
+def _human_readable_bytes(num_bytes: float) -> str:
+    for unit in ("B", "KB", "MB", "GB", "TB"):
+        if abs(num_bytes) < 1000.0:
+            return f"{num_bytes:3.1f} {unit}"
+        num_bytes /= 1000.0
+    return f"{num_bytes:.1f} PB"

src/lightning/data/utilities/format.py

@@ -10,7 +10,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
+import json
 import os
 import sys
 from functools import partial
@@ -27,6 +27,7 @@
 from lightning.fabric import Fabric
 from lightning.pytorch.demos.boring_classes import RandomDataset
 from lightning_utilities.core.imports import RequirementCache
+from lightning_utilities.test.warning import no_warning_call
 from torch.utils.data import DataLoader, Dataset
 
 _PIL_AVAILABLE = RequirementCache("PIL")
@@ -242,3 +243,35 @@ def test_streaming_dataset(tmpdir, monkeypatch):
 
     dataloader = DataLoader(dataset, num_workers=2, batch_size=2)
     assert len(dataloader) == 408
+
+
+def test_create_oversized_chunk_single_item(tmp_path):
+    cache = Cache(str(tmp_path), chunk_bytes=700)
+    with pytest.warns(UserWarning, match="An item was larger than the target chunk size"):
+        cache[0] = np.random.randint(0, 10, size=(10000,), dtype=np.uint8)
+
+
+def test_create_undersized_and_oversized_chunk(tmp_path):
+    cache = Cache(str(tmp_path), chunk_bytes=9000)  # target: 9KB chunks
+    with no_warning_call(UserWarning):
+        cache[0] = np.random.randint(0, 10, size=(500,), dtype=np.uint8)  # will result in undersized chunk
+        cache[1] = np.random.randint(0, 10, size=(10000,), dtype=np.uint8)  # will result in oversized chunk
+    with pytest.warns(UserWarning, match="An item was larger than the target chunk size"):
+        cache[2] = np.random.randint(0, 10, size=(150,), dtype=np.uint8)
+    with no_warning_call(UserWarning):
+        cache[3] = np.random.randint(0, 10, size=(200,), dtype=np.uint8)
+
+    cache.done()
+    cache.merge()
+
+    assert len(os.listdir(tmp_path)) == 4  # 3 chunks + 1 index file
+    with open(tmp_path / "index.json") as file:
+        index = json.load(file)
+
+    chunks = index["chunks"]
+    assert chunks[0]["chunk_size"] == 1
+    assert chunks[0]["filename"] == "chunk-0-0.bin"
+    assert chunks[1]["chunk_size"] == 1
+    assert chunks[1]["filename"] == "chunk-0-1.bin"
+    assert chunks[2]["chunk_size"] == 2
+    assert chunks[2]["filename"] == "chunk-0-2.bin"

tests/tests_data/streaming/test_cache.py

@@ -0,0 +1,18 @@
+from lightning.data.utilities.format import _human_readable_bytes
+
+
+def test_human_readable_bytes():
+    assert _human_readable_bytes(0) == "0.0 B"
+    assert _human_readable_bytes(1) == "1.0 B"
+    assert _human_readable_bytes(999) == "999.0 B"
+    assert _human_readable_bytes(int(1e3)) == "1.0 KB"
+    assert _human_readable_bytes(int(1e3 + 1e2)) == "1.1 KB"
+    assert _human_readable_bytes(int(1e6)) == "1.0 MB"
+    assert _human_readable_bytes(int(1e6 + 2e5)) == "1.2 MB"
+    assert _human_readable_bytes(int(1e9)) == "1.0 GB"
+    assert _human_readable_bytes(int(1e9 + 3e8)) == "1.3 GB"
+    assert _human_readable_bytes(int(1e12)) == "1.0 TB"
+    assert _human_readable_bytes(int(1e12 + 4e11)) == "1.4 TB"
+    assert _human_readable_bytes(int(1e15)) == "1.0 PB"
+    assert _human_readable_bytes(int(1e15 + 5e14)) == "1.5 PB"
+    assert _human_readable_bytes(int(1e18)) == "1000.0 PB"