Multithreaded downloads

.github/workflows/ci.yml

@@ -54,7 +54,7 @@ jobs:
         if: ${{ matrix.os == 'ubuntu-latest' }}
         run: echo "installing pinned version of setuptools-scm to fix seqeval installation on 3.7" && pip install "setuptools-scm==6.4.2"
       - name: Install uv
-        run: pip install --upgrade uv
+        run: pip install uv==0.1.29
       - name: Install dependencies
         run: |
           uv pip install --system "datasets[tests,metrics-tests] @ ."
@@ -89,7 +89,7 @@ jobs:
       - name: Upgrade pip
         run: python -m pip install --upgrade pip
       - name: Install uv
-        run: pip install --upgrade uv
+        run: pip install uv==0.1.29
       - name: Install dependencies
         run: uv pip install --system "datasets[tests] @ ."
       - name: Test with pytest

src/datasets/data_files.py

@@ -544,8 +544,8 @@ def _get_single_origin_metadata(
 
 def _get_origin_metadata(
     data_files: List[str],
-    max_workers=64,
     download_config: Optional[DownloadConfig] = None,
+    max_workers: int = 16,
 ) -> Tuple[str]:
     return thread_map(
         partial(_get_single_origin_metadata, download_config=download_config),

src/datasets/download/download_config.py

@@ -59,6 +59,8 @@ class DownloadConfig:
             Key/value pairs to be passed on to the dataset file-system backend, if any.
         download_desc (`str`, *optional*):
             A description to be displayed alongside with the progress bar while downloading the files.
+        disable_tqdm (`bool`, defaults to `False`):
+            Whether to disable the individual files download progress bar
     """
 
     cache_dir: Optional[Union[str, Path]] = None
@@ -78,6 +80,7 @@ class DownloadConfig:
     ignore_url_params: bool = False
     storage_options: Dict[str, Any] = field(default_factory=dict)
     download_desc: Optional[str] = None
+    disable_tqdm: bool = False
 
     def __post_init__(self, use_auth_token):
         if use_auth_token != "deprecated":

src/datasets/download/download_manager.py

@@ -17,6 +17,7 @@
 
 import enum
 import io
+import multiprocessing
 import os
 import posixpath
 import tarfile
@@ -27,6 +28,10 @@
 from itertools import chain
 from typing import Callable, Dict, Generator, List, Optional, Tuple, Union
 
+import fsspec
+from fsspec.core import url_to_fs
+from tqdm.contrib.concurrent import thread_map
+
 from .. import config
 from ..utils import tqdm as hf_tqdm
 from ..utils.deprecation_utils import DeprecatedEnum, deprecated
@@ -39,7 +44,7 @@
     url_or_path_join,
 )
 from ..utils.info_utils import get_size_checksum_dict
-from ..utils.logging import get_logger
+from ..utils.logging import get_logger, tqdm
 from ..utils.py_utils import NestedDataStructure, map_nested, size_str
 from ..utils.track import TrackedIterable, tracked_str
 from .download_config import DownloadConfig
@@ -427,7 +432,7 @@ def download(self, url_or_urls):
         if download_config.download_desc is None:
             download_config.download_desc = "Downloading data"
 
-        download_func = partial(self._download, download_config=download_config)
+        download_func = partial(self._download_batched, download_config=download_config)
 
         start_time = datetime.now()
         with stack_multiprocessing_download_progress_bars():
@@ -437,6 +442,8 @@ def download(self, url_or_urls):
                 map_tuple=True,
                 num_proc=download_config.num_proc,
                 desc="Downloading data files",
+                batched=True,
+                batch_size=-1,
             )
         duration = datetime.now() - start_time
         logger.info(f"Downloading took {duration.total_seconds() // 60} min")
@@ -451,6 +458,43 @@ def download(self, url_or_urls):
 
         return downloaded_path_or_paths.data
 
+    def _download_batched(
+        self,
+        url_or_filenames: List[str],
+        download_config: DownloadConfig,
+    ) -> List[str]:
+        if len(url_or_filenames) >= 16:
+            download_config = download_config.copy()
+            download_config.disable_tqdm = True
+            download_func = partial(self._download, download_config=download_config)
+
+            fs: fsspec.AbstractFileSystem
+            fs, path = url_to_fs(url_or_filenames[0], **download_config.storage_options)
+            size = 0
+            try:
+                size = fs.info(path).get("size", 0)
+            except Exception:
+                pass
+            max_workers = 16 if size < (20 << 20) else 1  # enable multithreading if files are small
+
+            return thread_map(
+                download_func,
+                url_or_filenames,
+                desc=download_config.download_desc or "Downloading",
+                unit="files",
+                position=multiprocessing.current_process()._identity[-1]  # contains the ranks of subprocesses
+                if os.environ.get("HF_DATASETS_STACK_MULTIPROCESSING_DOWNLOAD_PROGRESS_BARS") == "1"
+                and multiprocessing.current_process()._identity
+                else None,
+                max_workers=max_workers,
+                tqdm_class=tqdm,
+            )
+        else:
+            return [
+                self._download(url_or_filename, download_config=download_config)
+                for url_or_filename in url_or_filenames
+            ]
+
     def _download(self, url_or_filename: str, download_config: DownloadConfig) -> str:
         url_or_filename = str(url_or_filename)
         if is_relative_path(url_or_filename):

src/datasets/parallel/parallel.py

@@ -14,7 +14,7 @@ class ParallelBackendConfig:
 
 
 @experimental
-def parallel_map(function, iterable, num_proc, types, disable_tqdm, desc, single_map_nested_func):
+def parallel_map(function, iterable, num_proc, batched, batch_size, types, disable_tqdm, desc, single_map_nested_func):
     """
     **Experimental.** Apply a function to iterable elements in parallel, where the implementation uses either
     multiprocessing.Pool or joblib for parallelization.
@@ -32,21 +32,25 @@ def parallel_map(function, iterable, num_proc, types, disable_tqdm, desc, single
     """
     if ParallelBackendConfig.backend_name is None:
         return _map_with_multiprocessing_pool(
-            function, iterable, num_proc, types, disable_tqdm, desc, single_map_nested_func
+            function, iterable, num_proc, batched, batch_size, types, disable_tqdm, desc, single_map_nested_func
         )
 
-    return _map_with_joblib(function, iterable, num_proc, types, disable_tqdm, desc, single_map_nested_func)
+    return _map_with_joblib(
+        function, iterable, num_proc, batched, batch_size, types, disable_tqdm, desc, single_map_nested_func
+    )
 
 
-def _map_with_multiprocessing_pool(function, iterable, num_proc, types, disable_tqdm, desc, single_map_nested_func):
+def _map_with_multiprocessing_pool(
+    function, iterable, num_proc, batched, batch_size, types, disable_tqdm, desc, single_map_nested_func
+):
     num_proc = num_proc if num_proc <= len(iterable) else len(iterable)
     split_kwds = []  # We organize the splits ourselve (contiguous splits)
     for index in range(num_proc):
         div = len(iterable) // num_proc
         mod = len(iterable) % num_proc
         start = div * index + min(index, mod)
         end = start + div + (1 if index < mod else 0)
-        split_kwds.append((function, iterable[start:end], types, index, disable_tqdm, desc))
+        split_kwds.append((function, iterable[start:end], batched, batch_size, types, index, disable_tqdm, desc))
 
     if len(iterable) != sum(len(i[1]) for i in split_kwds):
         raise ValueError(
@@ -70,14 +74,17 @@ def _map_with_multiprocessing_pool(function, iterable, num_proc, types, disable_
     return mapped
 
 
-def _map_with_joblib(function, iterable, num_proc, types, disable_tqdm, desc, single_map_nested_func):
+def _map_with_joblib(
+    function, iterable, num_proc, batched, batch_size, types, disable_tqdm, desc, single_map_nested_func
+):
     # progress bar is not yet supported for _map_with_joblib, because tqdm couldn't accurately be applied to joblib,
     # and it requires monkey-patching joblib internal classes which is subject to change
     import joblib
 
     with joblib.parallel_backend(ParallelBackendConfig.backend_name, n_jobs=num_proc):
         return joblib.Parallel()(
-            joblib.delayed(single_map_nested_func)((function, obj, types, None, True, None)) for obj in iterable
+            joblib.delayed(single_map_nested_func)((function, obj, batched, batch_size, types, None, True, None))
+            for obj in iterable
         )
 
 

src/datasets/utils/file_utils.py

@@ -201,6 +201,7 @@ def cached_path(
             ignore_url_params=download_config.ignore_url_params,
             storage_options=download_config.storage_options,
             download_desc=download_config.download_desc,
+            disable_tqdm=download_config.disable_tqdm,
         )
     elif os.path.exists(url_or_filename):
         # File, and it exists.
@@ -335,7 +336,7 @@ def __init__(self, tqdm_kwargs=None, *args, **kwargs):
             super().__init__(tqdm_kwargs, *args, **kwargs)
 
 
-def fsspec_get(url, temp_file, storage_options=None, desc=None):
+def fsspec_get(url, temp_file, storage_options=None, desc=None, disable_tqdm=False):
     _raise_if_offline_mode_is_enabled(f"Tried to reach {url}")
     fs, path = url_to_fs(url, **(storage_options or {}))
     callback = TqdmCallback(
@@ -347,6 +348,7 @@ def fsspec_get(url, temp_file, storage_options=None, desc=None):
             if os.environ.get("HF_DATASETS_STACK_MULTIPROCESSING_DOWNLOAD_PROGRESS_BARS") == "1"
             and multiprocessing.current_process()._identity
             else None,
+            "disable": disable_tqdm,
         }
     )
     fs.get_file(path, temp_file.name, callback=callback)
@@ -373,7 +375,16 @@ def ftp_get(url, temp_file, timeout=10.0):
 
 
 def http_get(
-    url, temp_file, proxies=None, resume_size=0, headers=None, cookies=None, timeout=100.0, max_retries=0, desc=None
+    url,
+    temp_file,
+    proxies=None,
+    resume_size=0,
+    headers=None,
+    cookies=None,
+    timeout=100.0,
+    max_retries=0,
+    desc=None,
+    disable_tqdm=False,
 ) -> Optional[requests.Response]:
     headers = dict(headers) if headers is not None else {}
     headers["user-agent"] = get_datasets_user_agent(user_agent=headers.get("user-agent"))
@@ -405,6 +416,7 @@ def http_get(
         if os.environ.get("HF_DATASETS_STACK_MULTIPROCESSING_DOWNLOAD_PROGRESS_BARS") == "1"
         and multiprocessing.current_process()._identity
         else None,
+        disable=disable_tqdm,
     ) as progress:
         for chunk in response.iter_content(chunk_size=1024):
             progress.update(len(chunk))
@@ -464,6 +476,7 @@ def get_from_cache(
     ignore_url_params=False,
     storage_options=None,
     download_desc=None,
+    disable_tqdm=False,
 ) -> str:
     """
     Given a URL, look for the corresponding file in the local cache.
@@ -629,7 +642,9 @@ def temp_file_manager(mode="w+b"):
             if scheme == "ftp":
                 ftp_get(url, temp_file)
             elif scheme not in ("http", "https"):
-                fsspec_get(url, temp_file, storage_options=storage_options, desc=download_desc)
+                fsspec_get(
+                    url, temp_file, storage_options=storage_options, desc=download_desc, disable_tqdm=disable_tqdm
+                )
             else:
                 http_get(
                     url,
@@ -640,6 +655,7 @@ def temp_file_manager(mode="w+b"):
                     cookies=cookies,
                     max_retries=max_retries,
                     desc=download_desc,
+                    disable_tqdm=disable_tqdm,
                 )
 
         logger.info(f"storing {url} in cache at {cache_path}")