huggingface-hub 1.0.0rc6__py3-none-any.whl → 1.0.0rc7__py3-none-any.whl

huggingface_hub/hf_file_system.py

@@ -102,18 +102,22 @@ class HfFileSystem(fsspec.AbstractFileSystem):
         *args,
         endpoint: Optional[str] = None,
         token: Union[bool, str, None] = None,
+        block_size: Optional[int] = None,
         **storage_options,
     ):
         super().__init__(*args, **storage_options)
         self.endpoint = endpoint or constants.ENDPOINT
         self.token = token
         self._api = HfApi(endpoint=endpoint, token=token)
+        self.block_size = block_size
         # Maps (repo_type, repo_id, revision) to a 2-tuple with:
         #  * the 1st element indicating whether the repositoy and the revision exist
         #  * the 2nd element being the exception raised if the repository or revision doesn't exist
         self._repo_and_revision_exists_cache: dict[
             tuple[str, str, Optional[str]], tuple[bool, Optional[Exception]]
         ] = {}
+        # Maps parent directory path to path infos
+        self.dircache: dict[str, list[dict[str, Any]]] = {}
 
     def _repo_and_revision_exist(
         self, repo_type: str, repo_id: str, revision: Optional[str]
@@ -265,12 +269,15 @@ class HfFileSystem(fsspec.AbstractFileSystem):
         block_size: Optional[int] = None,
         **kwargs,
     ) -> "HfFileSystemFile":
+        block_size = block_size if block_size is not None else self.block_size
+        if block_size is not None:
+            kwargs["block_size"] = block_size
         if "a" in mode:
             raise NotImplementedError("Appending to remote files is not yet supported.")
         if block_size == 0:
-            return HfFileSystemStreamFile(self, path, mode=mode, revision=revision, block_size=block_size, **kwargs)
+            return HfFileSystemStreamFile(self, path, mode=mode, revision=revision, **kwargs)
         else:
-            return HfFileSystemFile(self, path, mode=mode, revision=revision, block_size=block_size, **kwargs)
+            return HfFileSystemFile(self, path, mode=mode, revision=revision, **kwargs)
 
     def _rm(self, path: str, revision: Optional[str] = None, **kwargs) -> None:
         resolved_path = self.resolve_path(path, revision=revision)
@@ -439,7 +446,7 @@ class HfFileSystem(fsspec.AbstractFileSystem):
                     common_path_depth = common_path[len(path) :].count("/")
                     maxdepth -= common_path_depth
                 out = [o for o in out if not o["name"].startswith(common_path + "/")]
-                for cached_path in self.dircache:
+                for cached_path in list(self.dircache):
                     if cached_path.startswith(common_path + "/"):
                         self.dircache.pop(cached_path, None)
                 self.dircache.pop(common_path, None)
@@ -923,6 +930,18 @@ class HfFileSystem(fsspec.AbstractFileSystem):
         # See https://github.com/huggingface/huggingface_hub/issues/1733
         raise NotImplementedError("Transactional commits are not supported.")
 
+    def __reduce__(self):
+        # re-populate the instance cache at HfFileSystem._cache and re-populate the cache attributes of every instance
+        return make_instance, (
+            type(self),
+            self.storage_args,
+            self.storage_options,
+            {
+                "dircache": self.dircache,
+                "_repo_and_revision_exists_cache": self._repo_and_revision_exists_cache,
+            },
+        )
+
 
 class HfFileSystemFile(fsspec.spec.AbstractBufferedFile):
     def __init__(self, fs: HfFileSystem, path: str, revision: Optional[str] = None, **kwargs):
@@ -986,9 +1005,8 @@ class HfFileSystemFile(fsspec.spec.AbstractBufferedFile):
     def read(self, length=-1):
         """Read remote file.
 
-        If `length` is not provided or is -1, the entire file is downloaded and read. On POSIX systems and if
-        `hf_transfer` is not enabled, the file is loaded in memory directly. Otherwise, the file is downloaded to a
-        temporary file and read from there.
+        If `length` is not provided or is -1, the entire file is downloaded and read. On POSIX systems the file is
+        loaded in memory directly. Otherwise, the file is downloaded to a temporary file and read from there.
         """
         if self.mode == "rb" and (length is None or length == -1) and self.loc == 0:
             with self.fs.open(self.path, "rb", block_size=0) as f:  # block_size=0 enables fast streaming
@@ -1158,3 +1176,10 @@ def _partial_read(response: httpx.Response, length: int = -1) -> bytes:
             return bytes(buf[:length])
 
     return bytes(buf)  # may be < length if response ended
+
+
+def make_instance(cls, args, kwargs, instance_cache_attributes_dict):
+    fs = cls(*args, **kwargs)
+    for attr, cached_value in instance_cache_attributes_dict.items():
+        setattr(fs, attr, cached_value)
+    return fs

huggingface_hub/inference/_client.py

@@ -135,7 +135,7 @@ class InferenceClient:
             Note: for better compatibility with OpenAI's client, `model` has been aliased as `base_url`. Those 2
             arguments are mutually exclusive. If a URL is passed as `model` or `base_url` for chat completion, the `(/v1)/chat/completions` suffix path will be appended to the URL.
         provider (`str`, *optional*):
-            Name of the provider to use for inference. Can be `"black-forest-labs"`, `"cerebras"`, `"cohere"`, `"fal-ai"`, `"featherless-ai"`, `"fireworks-ai"`, `"groq"`, `"hf-inference"`, `"hyperbolic"`, `"nebius"`, `"novita"`, `"nscale"`, `"openai"`, `publicai`, `"replicate"`, `"sambanova"`, `"scaleway"`, `"together"` or `"zai-org"`.
+            Name of the provider to use for inference. Can be `"black-forest-labs"`, `"cerebras"`, `"clarifai"`, `"cohere"`, `"fal-ai"`, `"featherless-ai"`, `"fireworks-ai"`, `"groq"`, `"hf-inference"`, `"hyperbolic"`, `"nebius"`, `"novita"`, `"nscale"`, `"openai"`, `publicai`, `"replicate"`, `"sambanova"`, `"scaleway"`, `"together"` or `"zai-org"`.
             Defaults to "auto" i.e. the first of the providers available for the model, sorted by the user's order in https://hf.co/settings/inference-providers.
             If model is a URL or `base_url` is passed, then `provider` is not used.
         token (`str`, *optional*):

huggingface_hub/inference/_generated/_async_client.py

@@ -126,7 +126,7 @@ class AsyncInferenceClient:
             Note: for better compatibility with OpenAI's client, `model` has been aliased as `base_url`. Those 2
             arguments are mutually exclusive. If a URL is passed as `model` or `base_url` for chat completion, the `(/v1)/chat/completions` suffix path will be appended to the URL.
         provider (`str`, *optional*):
-            Name of the provider to use for inference. Can be `"black-forest-labs"`, `"cerebras"`, `"cohere"`, `"fal-ai"`, `"featherless-ai"`, `"fireworks-ai"`, `"groq"`, `"hf-inference"`, `"hyperbolic"`, `"nebius"`, `"novita"`, `"nscale"`, `"openai"`, `publicai`, `"replicate"`, `"sambanova"`, `"scaleway"`, `"together"` or `"zai-org"`.
+            Name of the provider to use for inference. Can be `"black-forest-labs"`, `"cerebras"`, `"clarifai"`, `"cohere"`, `"fal-ai"`, `"featherless-ai"`, `"fireworks-ai"`, `"groq"`, `"hf-inference"`, `"hyperbolic"`, `"nebius"`, `"novita"`, `"nscale"`, `"openai"`, `publicai`, `"replicate"`, `"sambanova"`, `"scaleway"`, `"together"` or `"zai-org"`.
             Defaults to "auto" i.e. the first of the providers available for the model, sorted by the user's order in https://hf.co/settings/inference-providers.
             If model is a URL or `base_url` is passed, then `provider` is not used.
         token (`str`, *optional*):

huggingface_hub/inference/_providers/__init__.py

@@ -6,9 +6,10 @@ from huggingface_hub.inference._providers.featherless_ai import (
 )
 from huggingface_hub.utils import logging
 
-from ._common import TaskProviderHelper, _fetch_inference_provider_mapping
+from ._common import AutoRouterConversationalTask, TaskProviderHelper, _fetch_inference_provider_mapping
 from .black_forest_labs import BlackForestLabsTextToImageTask
 from .cerebras import CerebrasConversationalTask
+from .clarifai import ClarifaiConversationalTask
 from .cohere import CohereConversationalTask
 from .fal_ai import (
     FalAIAutomaticSpeechRecognitionTask,
@@ -50,6 +51,7 @@ logger = logging.get_logger(__name__)
 PROVIDER_T = Literal[
     "black-forest-labs",
     "cerebras",
+    "clarifai",
     "cohere",
     "fal-ai",
     "featherless-ai",
@@ -71,6 +73,8 @@ PROVIDER_T = Literal[
 
 PROVIDER_OR_POLICY_T = Union[PROVIDER_T, Literal["auto"]]
 
+CONVERSATIONAL_AUTO_ROUTER = AutoRouterConversationalTask()
+
 PROVIDERS: dict[PROVIDER_T, dict[str, TaskProviderHelper]] = {
     "black-forest-labs": {
         "text-to-image": BlackForestLabsTextToImageTask(),
@@ -78,6 +82,9 @@ PROVIDERS: dict[PROVIDER_T, dict[str, TaskProviderHelper]] = {
     "cerebras": {
         "conversational": CerebrasConversationalTask(),
     },
+    "clarifai": {
+        "conversational": ClarifaiConversationalTask(),
+    },
     "cohere": {
         "conversational": CohereConversationalTask(),
     },
@@ -201,13 +208,19 @@ def get_provider_helper(
 
     if provider is None:
         logger.info(
-            "Defaulting to 'auto' which will select the first provider available for the model, sorted by the user's order in https://hf.co/settings/inference-providers."
+            "No provider specified for task `conversational`. Defaulting to server-side auto routing."
+            if task == "conversational"
+            else "Defaulting to 'auto' which will select the first provider available for the model, sorted by the user's order in https://hf.co/settings/inference-providers."
         )
         provider = "auto"
 
     if provider == "auto":
         if model is None:
             raise ValueError("Specifying a model is required when provider is 'auto'")
+        if task == "conversational":
+            # Special case: we have a dedicated auto-router for conversational models. No need to fetch provider mapping.
+            return CONVERSATIONAL_AUTO_ROUTER
+
         provider_mapping = _fetch_inference_provider_mapping(model)
         provider = next(iter(provider_mapping)).provider
 

huggingface_hub/inference/_providers/_common.py

@@ -24,6 +24,7 @@ HARDCODED_MODEL_INFERENCE_MAPPING: dict[str, dict[str, InferenceProviderMapping]
     #                                    status="live")
     "cerebras": {},
     "cohere": {},
+    "clarifai": {},
     "fal-ai": {},
     "fireworks-ai": {},
     "groq": {},
@@ -278,6 +279,44 @@ class BaseConversationalTask(TaskProviderHelper):
         return filter_none({"messages": inputs, **parameters, "model": provider_mapping_info.provider_id})
 
 
+class AutoRouterConversationalTask(BaseConversationalTask):
+    """
+    Auto-router for conversational tasks.
+
+    We let the Hugging Face router select the best provider for the model, based on availability and user preferences.
+    This is a special case since the selection is done server-side (avoid 1 API call to fetch provider mapping).
+    """
+
+    def __init__(self):
+        super().__init__(provider="auto", base_url="https://router.huggingface.co")
+
+    def _prepare_base_url(self, api_key: str) -> str:
+        """Return the base URL to use for the request.
+
+        Usually not overwritten in subclasses."""
+        # Route to the proxy if the api_key is a HF TOKEN
+        if not api_key.startswith("hf_"):
+            raise ValueError("Cannot select auto-router when using non-Hugging Face API key.")
+        else:
+            return self.base_url  # No `/auto` suffix in the URL
+
+    def _prepare_mapping_info(self, model: Optional[str]) -> InferenceProviderMapping:
+        """
+        In auto-router, we don't need to fetch provider mapping info.
+        We just return a dummy mapping info with provider_id set to the HF model ID.
+        """
+        if model is None:
+            raise ValueError("Please provide an HF model ID.")
+
+        return InferenceProviderMapping(
+            provider="auto",
+            hf_model_id=model,
+            providerId=model,
+            status="live",
+            task="conversational",
+        )
+
+
 class BaseTextGenerationTask(TaskProviderHelper):
     """
     Base class for text-generation (completion) tasks.

huggingface_hub/inference/_providers/clarifai.py

@@ -0,0 +1,13 @@
+from ._common import BaseConversationalTask
+
+
+_PROVIDER = "clarifai"
+_BASE_URL = "https://api.clarifai.com"
+
+
+class ClarifaiConversationalTask(BaseConversationalTask):
+    def __init__(self):
+        super().__init__(provider=_PROVIDER, base_url=_BASE_URL)
+
+    def _prepare_route(self, mapped_model: str, api_key: str) -> str:
+        return "/v2/ext/openai/v1/chat/completions"

huggingface_hub/lfs.py

@@ -16,11 +16,9 @@
 
 import io
 import re
-import warnings
 from dataclasses import dataclass
 from math import ceil
 from os.path import getsize
-from pathlib import Path
 from typing import TYPE_CHECKING, BinaryIO, Iterable, Optional, TypedDict
 from urllib.parse import unquote
 
@@ -33,12 +31,10 @@ from .utils import (
     hf_raise_for_status,
     http_backoff,
     logging,
-    tqdm,
     validate_hf_hub_args,
 )
 from .utils._lfs import SliceFileObj
 from .utils.sha import sha256, sha_fileobj
-from .utils.tqdm import is_tqdm_disabled
 
 
 if TYPE_CHECKING:
@@ -332,23 +328,9 @@ def _upload_multi_part(operation: "CommitOperationAdd", header: dict, chunk_size
     # 1. Get upload URLs for each part
     sorted_parts_urls = _get_sorted_parts_urls(header=header, upload_info=operation.upload_info, chunk_size=chunk_size)
 
-    # 2. Upload parts (either with hf_transfer or in pure Python)
-    use_hf_transfer = constants.HF_HUB_ENABLE_HF_TRANSFER
-    if (
-        constants.HF_HUB_ENABLE_HF_TRANSFER
-        and not isinstance(operation.path_or_fileobj, str)
-        and not isinstance(operation.path_or_fileobj, Path)
-    ):
-        warnings.warn(
-            "hf_transfer is enabled but does not support uploading from bytes or BinaryIO, falling back to regular"
-            " upload"
-        )
-        use_hf_transfer = False
-
-    response_headers = (
-        _upload_parts_hf_transfer(operation=operation, sorted_parts_urls=sorted_parts_urls, chunk_size=chunk_size)
-        if use_hf_transfer
-        else _upload_parts_iteratively(operation=operation, sorted_parts_urls=sorted_parts_urls, chunk_size=chunk_size)
+    # 2. Upload parts (pure Python)
+    response_headers = _upload_parts_iteratively(
+        operation=operation, sorted_parts_urls=sorted_parts_urls, chunk_size=chunk_size
     )
 
     # 3. Send completion request
@@ -409,47 +391,3 @@ def _upload_parts_iteratively(
                 hf_raise_for_status(part_upload_res)
                 headers.append(part_upload_res.headers)
     return headers  # type: ignore
-
-
-def _upload_parts_hf_transfer(
-    operation: "CommitOperationAdd", sorted_parts_urls: list[str], chunk_size: int
-) -> list[dict]:
-    # Upload file using an external Rust-based package. Upload is faster but support less features (no progress bars).
-    try:
-        from hf_transfer import multipart_upload
-    except ImportError:
-        raise ValueError(
-            "Fast uploading using 'hf_transfer' is enabled (HF_HUB_ENABLE_HF_TRANSFER=1) but 'hf_transfer' package is"
-            " not available in your environment. Try `pip install hf_transfer`."
-        )
-
-    total = operation.upload_info.size
-    desc = operation.path_in_repo
-    if len(desc) > 40:
-        desc = f"(…){desc[-40:]}"
-
-    with tqdm(
-        unit="B",
-        unit_scale=True,
-        total=total,
-        initial=0,
-        desc=desc,
-        disable=is_tqdm_disabled(logger.getEffectiveLevel()),
-        name="huggingface_hub.lfs_upload",
-    ) as progress:
-        try:
-            output = multipart_upload(
-                file_path=operation.path_or_fileobj,
-                parts_urls=sorted_parts_urls,
-                chunk_size=chunk_size,
-                max_files=128,
-                parallel_failures=127,  # could be removed
-                max_retries=5,
-                callback=progress.update,
-            )
-        except Exception as e:
-            raise RuntimeError(
-                "An error occurred while uploading using `hf_transfer`. Consider disabling HF_HUB_ENABLE_HF_TRANSFER for"
-                " better error handling."
-            ) from e
-        return output

huggingface_hub/serialization/_torch.py

@@ -266,7 +266,7 @@ def save_torch_state_dict(
     safe_file_kwargs = {"metadata": per_file_metadata} if safe_serialization else {}
     for filename, tensors in state_dict_split.filename_to_tensors.items():
         shard = {tensor: state_dict[tensor] for tensor in tensors}
-        save_file_fn(shard, os.path.join(save_directory, filename), **safe_file_kwargs)
+        save_file_fn(shard, os.path.join(save_directory, filename), **safe_file_kwargs)  # ty: ignore[invalid-argument-type]
         logger.debug(f"Shard saved to {filename}")
 
     # Save the index (if any)

huggingface_hub/utils/__init__.py

@@ -75,7 +75,6 @@ from ._runtime import (
     get_gradio_version,
     get_graphviz_version,
     get_hf_hub_version,
-    get_hf_transfer_version,
     get_jinja_version,
     get_numpy_version,
     get_pillow_version,
@@ -94,7 +93,6 @@ from ._runtime import (
     is_google_colab,
     is_gradio_available,
     is_graphviz_available,
-    is_hf_transfer_available,
     is_jinja_available,
     is_notebook,
     is_numpy_available,

huggingface_hub/utils/_cache_manager.py

@@ -16,7 +16,6 @@
 
 import os
 import shutil
-import time
 from collections import defaultdict
 from dataclasses import dataclass
 from pathlib import Path
@@ -26,6 +25,7 @@ from huggingface_hub.errors import CacheNotFound, CorruptedCacheException
 
 from ..constants import HF_HUB_CACHE
 from . import logging
+from ._parsing import format_timesince
 from ._terminal import tabulate
 
 
@@ -79,7 +79,7 @@ class CachedFileInfo:
 
         Example: "2 weeks ago".
         """
-        return _format_timesince(self.blob_last_accessed)
+        return format_timesince(self.blob_last_accessed)
 
     @property
     def blob_last_modified_str(self) -> str:
@@ -89,7 +89,7 @@ class CachedFileInfo:
 
         Example: "2 weeks ago".
         """
-        return _format_timesince(self.blob_last_modified)
+        return format_timesince(self.blob_last_modified)
 
     @property
     def size_on_disk_str(self) -> str:
@@ -153,7 +153,7 @@ class CachedRevisionInfo:
 
         Example: "2 weeks ago".
         """
-        return _format_timesince(self.last_modified)
+        return format_timesince(self.last_modified)
 
     @property
     def size_on_disk_str(self) -> str:
@@ -223,7 +223,7 @@ class CachedRepoInfo:
 
         Example: "2 weeks ago".
         """
-        return _format_timesince(self.last_accessed)
+        return format_timesince(self.last_accessed)
 
     @property
     def last_modified_str(self) -> str:
@@ -233,7 +233,7 @@ class CachedRepoInfo:
 
         Example: "2 weeks ago".
         """
-        return _format_timesince(self.last_modified)
+        return format_timesince(self.last_modified)
 
     @property
     def size_on_disk_str(self) -> str:
@@ -244,6 +244,11 @@ class CachedRepoInfo:
         """
         return _format_size(self.size_on_disk)
 
+    @property
+    def cache_id(self) -> str:
+        """Canonical `type/id` identifier used across cache tooling."""
+        return f"{self.repo_type}/{self.repo_id}"
+
     @property
     def refs(self) -> dict[str, CachedRevisionInfo]:
         """
@@ -607,15 +612,12 @@ def scan_cache_dir(cache_dir: Optional[Union[str, Path]] = None) -> HFCacheInfo:
 
     You can also print a detailed report directly from the `hf` command line using:
     ```text
-    > hf cache scan
-    REPO ID                     REPO TYPE SIZE ON DISK NB FILES REFS                LOCAL PATH
-    --------------------------- --------- ------------ -------- ------------------- -------------------------------------------------------------------------
-    glue                        dataset         116.3K       15 1.17.0, main, 2.4.0 /Users/lucain/.cache/huggingface/hub/datasets--glue
-    google/fleurs               dataset          64.9M        6 main, refs/pr/1     /Users/lucain/.cache/huggingface/hub/datasets--google--fleurs
-    Jean-Baptiste/camembert-ner model           441.0M        7 main                /Users/lucain/.cache/huggingface/hub/models--Jean-Baptiste--camembert-ner
-    bert-base-cased             model             1.9G       13 main                /Users/lucain/.cache/huggingface/hub/models--bert-base-cased
-    t5-base                     model            10.1K        3 main                /Users/lucain/.cache/huggingface/hub/models--t5-base
-    t5-small                    model           970.7M       11 refs/pr/1, main     /Users/lucain/.cache/huggingface/hub/models--t5-small
+    > hf cache ls
+    ID                          SIZE     LAST_ACCESSED LAST_MODIFIED REFS
+    --------------------------- -------- ------------- ------------- -----------
+    dataset/nyu-mll/glue          157.4M 2 days ago    2 days ago    main script
+    model/LiquidAI/LFM2-VL-1.6B     3.2G 4 days ago    4 days ago    main
+    model/microsoft/UserLM-8b      32.1G 4 days ago    4 days ago    main
 
     Done in 0.0s. Scanned 6 repo(s) for a total of 3.4G.
     Got 1 warning(s) while scanning. Use -vvv to print details.
@@ -816,33 +818,6 @@ def _format_size(num: int) -> str:
     return f"{num_f:.1f}Y"
 
 
-_TIMESINCE_CHUNKS = (
-    # Label, divider, max value
-    ("second", 1, 60),
-    ("minute", 60, 60),
-    ("hour", 60 * 60, 24),
-    ("day", 60 * 60 * 24, 6),
-    ("week", 60 * 60 * 24 * 7, 6),
-    ("month", 60 * 60 * 24 * 30, 11),
-    ("year", 60 * 60 * 24 * 365, None),
-)
-
-
-def _format_timesince(ts: float) -> str:
-    """Format timestamp in seconds into a human-readable string, relative to now.
-
-    Vaguely inspired by Django's `timesince` formatter.
-    """
-    delta = time.time() - ts
-    if delta < 20:
-        return "a few seconds ago"
-    for label, divider, max_value in _TIMESINCE_CHUNKS:  # noqa: B007
-        value = round(delta / divider)
-        if max_value is not None and value <= max_value:
-            break
-    return f"{value} {label}{'s' if value > 1 else ''} ago"
-
-
 def _try_delete_path(path: Path, path_type: str) -> None:
     """Try to delete a local file or folder.
 

huggingface_hub/utils/_http.py

@@ -109,6 +109,20 @@ async def async_hf_request_event_hook(request: httpx.Request) -> None:
     return hf_request_event_hook(request)
 
 
+async def async_hf_response_event_hook(response: httpx.Response) -> None:
+    if response.status_code >= 400:
+        # If response will raise, read content from stream to have it available when raising the exception
+        # If content-length is not set or is too large, skip reading the content to avoid OOM
+        if "Content-length" in response.headers:
+            try:
+                length = int(response.headers["Content-length"])
+            except ValueError:
+                return
+
+            if length < 1_000_000:
+                await response.aread()
+
+
 def default_client_factory() -> httpx.Client:
     """
     Factory function to create a `httpx.Client` with the default transport.
@@ -125,7 +139,7 @@ def default_async_client_factory() -> httpx.AsyncClient:
     Factory function to create a `httpx.AsyncClient` with the default transport.
     """
     return httpx.AsyncClient(
-        event_hooks={"request": [async_hf_request_event_hook]},
+        event_hooks={"request": [async_hf_request_event_hook], "response": [async_hf_response_event_hook]},
         follow_redirects=True,
         timeout=httpx.Timeout(constants.DEFAULT_REQUEST_TIMEOUT, write=60.0),
     )
@@ -626,8 +640,16 @@ def _format(error_type: type[HfHubHTTPError], custom_message: str, response: htt
         try:
             data = response.json()
         except httpx.ResponseNotRead:
-            response.read()  # In case of streaming response, we need to read the response first
-            data = response.json()
+            try:
+                response.read()  # In case of streaming response, we need to read the response first
+                data = response.json()
+            except RuntimeError:
+                # In case of async streaming response, we can't read the stream here.
+                # In practice if user is using the default async client from `get_async_client`, the stream will have
+                # already been read in the async event hook `async_hf_response_event_hook`.
+                #
+                # Here, we are skipping reading the response to avoid RuntimeError but it happens only if async + stream + used httpx.AsyncClient directly.
+                data = {}
 
         error = data.get("error")
         if error is not None:

huggingface_hub/utils/_parsing.py

@@ -0,0 +1,98 @@
+# coding=utf-8
+# Copyright 2025-present, the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+"""Parsing helpers shared across modules."""
+
+import re
+import time
+from typing import Dict
+
+
+RE_NUMBER_WITH_UNIT = re.compile(r"(\d+)([a-z]+)", re.IGNORECASE)
+
+BYTE_UNITS: Dict[str, int] = {
+    "k": 1_000,
+    "m": 1_000_000,
+    "g": 1_000_000_000,
+    "t": 1_000_000_000_000,
+    "p": 1_000_000_000_000_000,
+}
+
+TIME_UNITS: Dict[str, int] = {
+    "s": 1,
+    "m": 60,
+    "h": 60 * 60,
+    "d": 24 * 60 * 60,
+    "w": 7 * 24 * 60 * 60,
+    "mo": 30 * 24 * 60 * 60,
+    "y": 365 * 24 * 60 * 60,
+}
+
+
+def parse_size(value: str) -> int:
+    """Parse a size expressed as a string with digits and unit (like `"10MB"`) to an integer (in bytes)."""
+    return _parse_with_unit(value, BYTE_UNITS)
+
+
+def parse_duration(value: str) -> int:
+    """Parse a duration expressed as a string with digits and unit (like `"10s"`) to an integer (in seconds)."""
+    return _parse_with_unit(value, TIME_UNITS)
+
+
+def _parse_with_unit(value: str, units: Dict[str, int]) -> int:
+    """Parse a numeric value with optional unit."""
+    stripped = value.strip()
+    if not stripped:
+        raise ValueError("Value cannot be empty.")
+    try:
+        return int(value)
+    except ValueError:
+        pass
+
+    match = RE_NUMBER_WITH_UNIT.fullmatch(stripped)
+    if not match:
+        raise ValueError(f"Invalid value '{value}'. Must match pattern '\\d+[a-z]+' or be a plain number.")
+
+    number = int(match.group(1))
+    unit = match.group(2).lower()
+
+    if unit not in units:
+        raise ValueError(f"Unknown unit '{unit}'. Must be one of {list(units.keys())}.")
+
+    return number * units[unit]
+
+
+def format_timesince(ts: float) -> str:
+    """Format timestamp in seconds into a human-readable string, relative to now.
+
+    Vaguely inspired by Django's `timesince` formatter.
+    """
+    _TIMESINCE_CHUNKS = (
+        # Label, divider, max value
+        ("second", 1, 60),
+        ("minute", 60, 60),
+        ("hour", 60 * 60, 24),
+        ("day", 60 * 60 * 24, 6),
+        ("week", 60 * 60 * 24 * 7, 6),
+        ("month", 60 * 60 * 24 * 30, 11),
+        ("year", 60 * 60 * 24 * 365, None),
+    )
+    delta = time.time() - ts
+    if delta < 20:
+        return "a few seconds ago"
+    for label, divider, max_value in _TIMESINCE_CHUNKS:  # noqa: B007
+        value = round(delta / divider)
+        if max_value is not None and value <= max_value:
+            break
+    return f"{value} {label}{'s' if value > 1 else ''} ago"