unify data loading from HF and from disk

ghstack-source-id: 932e7cce828a15c788b34f07c264e119068777fe
Pull Request resolved: pytorch#287

.gitignore

@@ -10,6 +10,5 @@ dist/*
 data
 out
 wandb
-*.json
 
 torchtitan/datasets/**/*.model

torchtitan/datasets/hf_datasets.py

@@ -12,9 +12,11 @@
 from torchtitan.datasets.tokenizer import Tokenizer
 from torchtitan.logging_utils import logger
 
-from datasets import load_dataset, load_from_disk
+from datasets import load_dataset
 from datasets.distributed import split_dataset_by_node
 
+# map from dataset name to a local directory, or
+# a dataset repository on the HF hub
 _supported_datasets = {
     "c4_mini": "torchtitan/datasets/c4_mini",
     "c4": "allenai/c4",
@@ -66,7 +68,7 @@ def __init__(
         rank: int = 0,
         infinite: bool = False,
     ) -> None:
-        # allow user to pass in a local path to use unsupported datasets
+        # allow user to pass in a (local or HF hub) path to use unsupported datasets
         if dataset_name not in _supported_datasets:
             if dataset_path:
                 logger.warning(
@@ -79,32 +81,18 @@ def __init__(
                     f"Supported datasets are: {list(_supported_datasets.keys())}."
                 )
 
-        # special case to auto-load c4_mini (and any future datasets) from local dir
-        if dataset_name == "c4_mini":
-            dataset_path = f"torchtitan/datasets/{dataset_name}"
+        if not dataset_path:
+            dataset_path = _supported_datasets[dataset_name]
+        logger.info(f"Preparing {dataset_name} dataset from {dataset_path}")
 
-        # TODO: This is a temporary solution for small datasets.
-        #       For large datasets we need to use a more scalable approach,
-        #       and support shuffling and checkpointing.
-        if dataset_path:
-            logger.info(f"Loading {dataset_name} dataset locally from {dataset_path}")
-            ds = load_from_disk(dataset_path)
+        if dataset_name == "c4":
+            # c4 is huge, and requires both streaming and language selection
+            # (we default to en)
+            ds = load_dataset(dataset_path, name="en", split="train", streaming=True)
         else:
-            logger.info(f"Preparing {dataset_name} dataset from HuggingFace")
-            # Setting `streaming=True` works for large dataset, but is slightly
-            # slower and unstable.
-            if dataset_name == "c4":
-                # c4 is huge, and requires both streaming and language selection
-                # (we default to en).
-                ds = load_dataset(
-                    _supported_datasets[dataset_name],
-                    "en",
-                    split="train",
-                    streaming=True,
-                )
-            else:
-                ds = load_dataset(_supported_datasets[dataset_name], split="train")
+            ds = load_dataset(dataset_path, split="train")
 
+        # TODO: support shuffling and checkpointing
         self.dataset_name = dataset_name
         self._data = split_dataset_by_node(ds, rank, world_size)
         self._tokenizer = tokenizer