Allow users to customize dataloader

ghstack-source-id: c9999c58aab4c0148e3a9f882e19cdc7eb29b55d
Pull Request resolved: #836

tests/unit_tests/test_dataset_checkpointing.py

@@ -5,8 +5,7 @@
 # LICENSE file in the root directory of this source tree.
 
 import torch
-from torchtitan.datasets.hf_datasets import build_hf_data_loader
-from torchtitan.datasets.tokenizer import build_tokenizer
+from torchtitan.datasets.hf_datasets import build_hf_dataloader
 
 
 class TestDatasetCheckpointing:
@@ -41,13 +40,12 @@ def test_c4_resumption(self):
     def _build_dataloader(
         self, dataset_name, dataset_path, batch_size, seq_len, world_size, rank
     ):
-        tokenizer = build_tokenizer("tiktoken", "./tests/assets/test_tiktoken.model")
-        return build_hf_data_loader(
+        return build_hf_dataloader(
             dataset_name=dataset_name,
             dataset_path=dataset_path,
-            tokenizer=tokenizer,
+            tokenizer_path="./tests/assets/test_tiktoken.model",
             batch_size=1,
             seq_len=1024,
-            world_size=4,
-            rank=0,
+            dp_world_size=4,
+            dp_rank=0,
         )

tests/unit_tests/test_train_spec.py

@@ -10,6 +10,7 @@
 import torch
 import torch.nn as nn
 from torchtitan.config_manager import JobConfig
+from torchtitan.datasets import build_hf_dataloader
 from torchtitan.models.llama import parallelize_llama, pipeline_llama
 from torchtitan.optimizer import (
     build_lr_schedulers,
@@ -60,6 +61,7 @@ def test_register_train_spec(self):
             pipelining_fn=pipeline_llama,
             build_optimizers_fn=build_optimizers,
             build_lr_schedulers_fn=build_lr_schedulers,
+            build_dataloader_fn=build_hf_dataloader,
         )
         register_train_spec(spec)
         new_spec = get_train_spec("fake")
@@ -78,6 +80,7 @@ def test_optim_hook(self):
             pipelining_fn=pipeline_llama,
             build_optimizers_fn=fake_build_optimizers,
             build_lr_schedulers_fn=build_lr_schedulers,
+            build_dataloader_fn=build_hf_dataloader,
         )
         register_train_spec(spec)
         new_spec = get_train_spec("fake2")

torchtitan/dataloader.py

@@ -0,0 +1,98 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+#
+# Copyright (c) Meta Platforms, Inc. All Rights Reserved.
+
+import pickle
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from typing import Any, Callable, TypeAlias
+
+from torch.distributed.checkpoint.stateful import Stateful
+from torch.utils.data import IterableDataset
+from torchdata.stateful_dataloader import StatefulDataLoader
+
+from torchtitan.datasets.tokenizer import Tokenizer
+
+
+class BaseDataLoader(Stateful, ABC):
+    """Base class for all dataloaders.
+
+    This is used to enforce that all dataloaders have the methods defined in ``Stateful``,
+    ``state_dict()`` and ``load_state_dict()``.
+    """
+
+    @abstractmethod
+    def __iter__(self):
+        ...
+
+
+class ParallelAwareDataloader(StatefulDataLoader, BaseDataLoader):
+    """Dataloader that is aware of distributed data parallelism.
+
+    This dataloader is used to load data in a distributed data parallel fashion. It also
+    utilizes ``torchdata.stateful_dataloader.StatefulDataLoader`` to implement the necessary
+    methods such as ``__iter__``.
+
+    Args:
+        dataset (IterableDataset): The dataset to iterate over.
+        tokenizer (Tokenizer): The tokenizer to use to tokenize the dataset.
+        dp_rank: Data parallelism rank for this dataloader.
+        dp_world_size: The world size of the data parallelism.
+        batch_size: The batch size to use for each iteration.
+    """
+
+    tokenizer: Tokenizer
+    dp_rank: int
+    dp_world_size: int
+    batch_size: int
+
+    def __init__(
+        self,
+        dataset: IterableDataset,
+        tokenizer: Tokenizer,
+        dp_rank: int,
+        dp_world_size: int,
+        batch_size: int,
+    ):
+        self.dp_world_size = dp_world_size
+        self.dp_rank = dp_rank
+        self.batch_size = batch_size
+        self.tokenizer = tokenizer
+        super().__init__(dataset, batch_size)
+        self._rank_id = f"dp_rank_{dp_rank}"
+
+    def state_dict(self) -> dict[str, Any]:
+        # Store state only for dp rank to avoid replicating the same state across other dimensions.
+        return {
+            # We don't have to use pickle as DCP will serialize the state_dict. However,
+            # we have to keep this for backward compatibility.
+            self._rank_id: pickle.dumps(super().state_dict()),
+            "world_size": self.dp_world_size,
+        }
+
+    def load_state_dict(self, state_dict: dict[str, Any]) -> None:
+        # State being empty is valid.
+        if not state_dict:
+            return
+
+        if self._rank_id not in state_dict:
+            logger.warning(
+                f"DataLoader state is empty for dp rank {self.dp_rank}, "
+                "expected key {self._rank_id}"
+            )
+            return
+
+        assert self.dp_world_size == state_dict["world_size"], (
+            "dp_degree is inconsistent before and after checkpoint, "
+            "dataloader resharding is not supported yet."
+        )
+        # We don't have to use pickle as DCP will serialize the state_dict. However, we have to
+        # keep this for backward compatibility.
+        super().load_state_dict(pickle.loads(state_dict[self._rank_id]))
+
+
+DataLoaderBuilder: TypeAlias = Callable[[...], BaseDataLoader]

torchtitan/datasets/__init__.py

@@ -4,10 +4,10 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-from torchtitan.datasets.hf_datasets import build_hf_data_loader
+from torchtitan.datasets.hf_datasets import build_hf_dataloader
 from torchtitan.datasets.tokenizer import build_tokenizer
 
 __all__ = [
-    "build_hf_data_loader",
+    "build_hf_dataloader",
     "build_tokenizer",
 ]

torchtitan/datasets/hf_datasets.py

@@ -4,28 +4,28 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-import pickle
 from dataclasses import dataclass
-from typing import Any, Callable, Dict, List, Optional
+from typing import Any, Callable, Optional
 
 import torch
+
+from datasets import Dataset, load_dataset
+from datasets.distributed import split_dataset_by_node
 from torch.distributed.checkpoint.stateful import Stateful
 from torch.utils.data import IterableDataset
-from torchdata.stateful_dataloader import StatefulDataLoader
 
-from torchtitan.datasets.tokenizer import Tokenizer
-from torchtitan.logging import logger
+from torchtitan.dataloader import ParallelAwareDataloader
 
-from datasets import Dataset, load_dataset
-from datasets.distributed import split_dataset_by_node
+from torchtitan.datasets.tokenizer import build_tokenizer, Tokenizer
+from torchtitan.logging import logger
 
 
 def _load_c4_dataset(dataset_path: str):
     """Load C4 dataset with default configuration."""
     return load_dataset(dataset_path, name="en", split="train", streaming=True)
 
 
-def _process_c4_text(sample: Dict[str, Any]) -> str:
+def _process_c4_text(sample: dict[str, Any]) -> str:
     """Process C4 dataset sample text."""
     return sample["text"]
 
@@ -75,8 +75,8 @@ def __init__(
         dataset_path: Optional[str],
         tokenizer: Tokenizer,
         seq_len: int = 2048,
-        world_size: int = 1,
-        rank: int = 0,
+        dp_rank: int = 0,
+        dp_world_size: int = 1,
         infinite: bool = False,
     ) -> None:
         # Force lowercase for consistent comparison
@@ -88,15 +88,15 @@ def __init__(
         ds = dataset_loader(path)
 
         self.dataset_name = dataset_name
-        self._data = split_dataset_by_node(ds, rank, world_size)
+        self._data = split_dataset_by_node(ds, dp_rank, dp_world_size)
         self._tokenizer = tokenizer
         self.seq_len = seq_len
         self.infinite = infinite
         self._text_processor = text_processor
 
         # Variables for checkpointing
         self._sample_idx = 0
-        self._all_tokens: List[int] = []
+        self._all_tokens: list[int] = []
 
     def _get_data_iter(self):
         if self._sample_idx == 0:
@@ -142,56 +142,33 @@ def state_dict(self):
         return {"token_buffer": self._all_tokens, "sample_idx": self._sample_idx}
 
 
-class DPAwareDataLoader(StatefulDataLoader, Stateful):
-    """
-    A wrapper around the StatefulDataLoader that ensures that the state is stored only once per DP rank.
-    """
-
-    def __init__(
-        self, dp_rank: int, hf_ds: IterableDataset, batch_size: int, world_size: int
-    ):
-        super().__init__(hf_ds, batch_size)
-        self._dp_rank = dp_rank
-        self._rank_id = f"dp_rank_{dp_rank}"
-        # Data loader resharding is not yet supported, so we need to store the world size to compare during loading
-        # raise error if dp_word_size does not match.
-        self._world_size = world_size
-
-    def state_dict(self) -> Dict[str, Any]:
-        # Store state only for dp rank to avoid replicating the same state across other dimensions
-        return {
-            self._rank_id: pickle.dumps(super().state_dict()),
-            "world_size": self._world_size,
-        }
-
-    def load_state_dict(self, state_dict: Dict[str, Any]) -> None:
-        # State being empty is valid
-        if not state_dict:
-            return
-
-        if self._rank_id not in state_dict:
-            logger.warning(
-                f"DataLoader state is empty for dp rank {self._dp_rank}, expected key {self._rank_id}"
-            )
-            return
-        assert (
-            self._world_size == state_dict["world_size"]
-        ), "dp_degree is inconsistent before and after checkpoint, dataloader resharding is not supported yet."
-        super().load_state_dict(pickle.loads(state_dict[self._rank_id]))
-
-
-def build_hf_data_loader(
+def build_hf_dataloader(
     dataset_name: str,
     dataset_path: Optional[str],
-    tokenizer: Tokenizer,
+    tokenizer_path: str,
     batch_size: int,
     seq_len: int,
-    world_size: int,
-    rank: int,
+    dp_rank: int,
+    dp_world_size: int,
     infinite: bool = True,
-):
+) -> ParallelAwareDataloader:
     """Build a data loader for HuggingFace datasets."""
+    tokenizer = build_tokenizer("tiktoken", tokenizer_path)
+
     hf_ds = HuggingFaceDataset(
-        dataset_name, dataset_path, tokenizer, seq_len, world_size, rank, infinite
+        dataset_name=dataset_name,
+        dataset_path=dataset_path,
+        tokenizer=tokenizer,
+        seq_len=seq_len,
+        dp_rank=dp_rank,
+        dp_world_size=dp_world_size,
+        infinite=infinite,
+    )
+
+    return ParallelAwareDataloader(
+        dataset=hf_ds,
+        tokenizer=tokenizer,
+        dp_rank=dp_rank,
+        dp_world_size=dp_world_size,
+        batch_size=batch_size,
     )
-    return DPAwareDataLoader(rank, hf_ds, batch_size=batch_size, world_size=world_size)

torchtitan/models/__init__.py

@@ -8,6 +8,3 @@
 # Import the built-in models here so that the corresponding register_model_spec()
 # will be called.
 import torchtitan.models.llama  # noqa: F401
-
-
-model_name_to_tokenizer = {"llama3": "tiktoken"}

torchtitan/models/llama/__init__.py

@@ -6,6 +6,7 @@
 #
 # Copyright (c) Meta Platforms, Inc. All Rights Reserved.
 
+from torchtitan.datasets import build_hf_dataloader
 from torchtitan.models.llama.model import Transformer, TransformerModelArgs
 from torchtitan.optimizer import build_lr_schedulers, build_optimizers
 from torchtitan.train_spec import register_train_spec, TrainSpec
@@ -65,5 +66,6 @@
         pipelining_fn=pipeline_llama,
         build_optimizers_fn=build_optimizers,
         build_lr_schedulers_fn=build_lr_schedulers,
+        build_dataloader_fn=build_hf_dataloader,
     )
 )

torchtitan/train_spec.py

@@ -6,14 +6,14 @@
 #
 # Copyright (c) Meta Platforms, Inc. All Rights Reserved.
 
-
 from dataclasses import dataclass
-from typing import Callable, Dict, Protocol, Type, TypeAlias
+from typing import Callable, Protocol, Type, TypeAlias
 
 import torch.nn as nn
 from torch.distributed.pipelining.schedules import _PipelineSchedule
 
 from torchtitan.config_manager import JobConfig
+from torchtitan.dataloader import DataLoaderBuilder
 from torchtitan.optimizer import LRSchedulersContainer, OptimizersContainer
 
 
@@ -53,15 +53,14 @@ def from_model_args(args: BaseModelArgs) -> nn.Module:
 class TrainSpec:
     name: str
     cls: Type[nn.Module]
-    config: Dict[str, BaseModelArgs]
+    config: dict[str, BaseModelArgs]
     parallelize_fn: Callable[[nn.Module], None]
     pipelining_fn: Callable[
         [nn.Module], tuple[_PipelineSchedule, list[nn.Module], bool, bool]
     ]
     build_optimizers_fn: OptimizersBuilder
     build_lr_schedulers_fn: LRSchedulersBuilder
-
-    # TODO: Add a ``build_dataloader_fn``
+    build_dataloader_fn: DataLoaderBuilder
 
     # TODO: Add a FQN convert fn to allow users to load checkpoints from
     # HuggingFace or other sources that have different FQN conventions.