add support for s3 checkpoint: need a fix on check_path_is_local

src/nanotron/config/config.py

@@ -2,6 +2,7 @@
 import os
 from dataclasses import dataclass, fields
 from pathlib import Path
+from datasets.download.streaming_download_manager import xPath
 from typing import List, Optional, Type, Union
 
 import dacite
@@ -91,6 +92,22 @@ def __post_init__(self):
             self.hf_dataset_splits = "train"
 
 
+@dataclass
+class S3UploadArgs:
+    """Arguments related to uploading checkpoints on s3"""
+
+    upload_s3_path: xPath
+    remove_after_upload: bool
+    s5cmd_numworkers: Optional[int]
+    s5cmd_concurrency: Optional[int]
+    s5cmd_path: Optional[xPath]
+
+    def __post_init__(self):
+        if isinstance(self.upload_s3_path, str):
+            self.upload_s3_path = xPath(self.upload_s3_path)
+        if isinstance(self.s5cmd_path, str):
+            self.s5cmd_path = xPath(self.s5cmd_path)
+
 @dataclass
 class NanosetDatasetsArgs:
     dataset_folder: Union[str, dict, List[str]]
@@ -338,13 +355,18 @@ class Config:
     data_stages: Optional[List[DatasetStageArgs]] = None
     profiler: Optional[ProfilerArgs] = None
     lighteval: Optional[LightEvalConfig] = None
+    s3_upload : Optional[S3UploadArgs] = None
 
     @classmethod
     def create_empty(cls):
         cls_fields = fields(cls)
         return cls(**{f.name: None for f in cls_fields})
 
     def __post_init__(self):
+
+        if self.s3_upload is not None:
+            self.s3_upload.__post_init__()
+
         # Some final sanity checks across separate arguments sections:
         if self.profiler is not None and self.profiler.profiler_export_path is not None:
             assert self.tokens.train_steps < 10

src/nanotron/helpers.py

@@ -52,8 +52,9 @@ def _vocab_size_with_padding(orig_vocab_size: int, pg_size: int, make_vocab_size
 
     multiple = make_vocab_size_divisible_by * pg_size
     after = int(ceil(orig_vocab_size / multiple) * multiple)
-
+    print("hello")
     if after != orig_vocab_size:
+        print("i'm in")
         log_rank(
             f"[Vocab Size Padding] Padded vocab (size: {orig_vocab_size}) with {after - orig_vocab_size} dummy tokens (new size: {after})",
             logger=logger,

src/nanotron/s3_checkpoints/__init__.py

@@ -0,0 +1,4 @@
+from .fsspec import check_path_is_local, fs_copy, fs_open
+from .s3_mover import S3Mover
+
+__all__ = ["S3Mover", "fs_open", "fs_copy", "check_path_is_local"]

src/nanotron/s3_checkpoints/fsspec.py

@@ -0,0 +1,38 @@
+import contextlib
+from pathlib import Path
+from typing import Tuple, Union
+
+import fsspec
+from fsspec.implementations import local
+
+
+def get_filesystem_and_path(path: Path, storage_options=None) -> Tuple[fsspec.AbstractFileSystem, str]:
+    # Use supported filesystems in `fsspec`. If you need another one, please use `fsspec.registry.register_implementation`
+    # DO NOT USE `mode` argument as it adds a suffix `0.part` when using `mode="w"`.
+    fs, _, paths = fsspec.core.get_fs_token_paths(str(path), storage_options=storage_options)
+    assert len(paths) == 1
+    return fs, paths[0]
+
+
+@contextlib.contextmanager
+def fs_open(
+    file: Union[str, Path],
+    mode="r",
+):
+    # TODO @thomasw21: pass storage options
+    fs, path = get_filesystem_and_path(file)
+    with fs.open(path, mode=mode) as f:
+        yield f
+
+
+def fs_copy(
+    input_file: Union[str, Path],
+    output_file: Union[str, Path],
+):
+    """Copy file from input to output (possibly on s3/other fs)"""
+    with fs_open(input_file, mode="rb") as fi, fs_open(output_file, mode="wb") as fo:
+        fo.write(fi.read())
+
+
+def check_path_is_local(path: Path, storage_options=None) -> bool:
+    return isinstance(get_filesystem_and_path(path=path, storage_options=storage_options)[0], local.LocalFileSystem)