v0.23.0: LLMs with tools, seamless downloads, and much more!

📁 Seamless download to local dir

The 0.23.0 release comes with a big revamp of the download process, especially when it comes to downloading to a local directory. Previously the process was still involving the cache directory and symlinks which led to misconceptions and a suboptimal user experience. The new workflow involves a .cache/huggingface/ folder, similar to the .git/ one, that keeps track of the progress of a download. The main features are:

• no symlinks
• no local copy
• don't re-download when not necessary
• same behavior on both Unix and Windows
• unrelated to cache-system

Example to download q4 GGUF file for microsoft/Phi-3-mini-4k-instruct-gguf:

# Download q4 GGUF file from 
huggingface-cli download microsoft/Phi-3-mini-4k-instruct-gguf Phi-3-mini-4k-instruct-q4.gguf --local-dir=data/phi3

With this addition, interrupted downloads are now resumable! This applies both for downloads in local and cache directories which should greatly improve UX for users with slow/unreliable connections. In this regard, the resume_download parameter is now deprecated (not relevant anymore).

• Revamp download to local dir process by @Wauplin in #2223
• Rename .huggingface/ folder to .cache/huggingface/ by @Wauplin in #2262

💡 Grammar and Tools in InferenceClient

It is now possible to provide a list of tools when chatting with a model using the InferenceClient! This major improvement has been made possible thanks to TGI that handle them natively.

>>> from huggingface_hub import InferenceClient

# Ask for weather in the next days using tools
>>> client = InferenceClient("meta-llama/Meta-Llama-3-70B-Instruct")
>>> messages = [
...     {"role": "system", "content": "Don't make assumptions about what values to plug into functions. Ask for clarification if a user request is ambiguous."},
...     {"role": "user", "content": "What's the weather like the next 3 days in San Francisco, CA?"},
... ]
>>> tools = [
...     {
...         "type": "function",
...         "function": {
...             "name": "get_current_weather",
...             "description": "Get the current weather",
...             "parameters": {
...                 "type": "object",
...                 "properties": {
...                     "location": {
...                         "type": "string",
...                         "description": "The city and state, e.g. San Francisco, CA",
...                     },
...                     "format": {
...                         "type": "string",
...                         "enum": ["celsius", "fahrenheit"],
...                         "description": "The temperature unit to use. Infer this from the users location.",
...                     },
...                 },
...                 "required": ["location", "format"],
...             },
...         },
...     },
...     ...
... ]
>>> response = client.chat_completion(
...     model="meta-llama/Meta-Llama-3-70B-Instruct",
...     messages=messages,
...     tools=tools,
...     tool_choice="auto",
...     max_tokens=500,
... )
>>> response.choices[0].message.tool_calls[0].function
ChatCompletionOutputFunctionDefinition(
    arguments={
        'location': 'San Francisco, CA',
        'format': 'fahrenheit',
        'num_days': 3
    },
    name='get_n_day_weather_forecast',
    description=None
)

It is also possible to provide grammar rules to the text_generation task. This ensures that the output follows a precise JSON Schema specification or matches a regular expression. For more details about it, check out the Guidance guide from Text-Generation-Inference docs.

• Add support for Grammar/Tools + TGI-based specs in InferenceClient by @Wauplin in #2237

⚙️ Other

Mention more chat-completion task instead of conversation in documentation.

• Add chat_completion and remove conversational from Inference guide by @Wauplin in #2215

chat-completion relies on server-side rendering in all cases, including when model is transformers-backed. Previously it was only the case for TGI-backed models and templates were rendered client-side otherwise.

• Render chat-template server-side for transformers-backed models by @Wauplin in #2258

Improved logic to determine whether a model is served via TGI or transformers.

• Raise error in chat completion when unprocessable by @Wauplin in #2257

• Document more chat_completion by @Wauplin in #2260

🌐 📚 Korean community is on fire!

The PseudoLab team is a non-profit dedicated to make AI more accessible in the Korean-speaking community. In the past few weeks, their team of contributors managed to translated (almost) entirely the huggingface_hub documentation. Huge shout-out to the coordination on this task! Documentation can be accessed here.

• 🌐 [i18n-KO] Translated guides/webhooks_server.md to Korean by @nuatmochoi in #2145
• 🌐 [i18n-KO] Translated reference/login.md to Korean by @SeungAhSon in #2151
• 🌐 [i18n-KO] Translated package_reference/tensorboard.md to Korean by @fabxoe in #2173
• 🌐 [i18n-KO] Translated package_reference/inference_client.md to Korean by @cjfghk5697 in #2178
• 🌐 [i18n-KO] Translated reference/inference_endpoints.md to Korean by @harheem in #2180
• 🌐 [i18n-KO] Translated package_reference/file_download.md to Korean by @seoyoung-3060 in #2184
• 🌐 [i18n-KO] Translated package_reference/cache.md to Korean by @nuatmochoi in #2191
• 🌐 [i18n-KO] Translated package_reference/collections.md to Korean by @boyunJang in #2214
• 🌐 [i18n-KO] Translated package_reference/inference_types.md to Korean by @fabxoe in #2171
• 🌐 [i18n-KO] Translated guides/upload.md to Korean by @junejae in #2139
• 🌐 [i18n-KO] Translated reference/repository.md to Korean by @junejae in #2189
• 🌐 [i18n-KO] Translated package_reference/space_runtime.md to Korean by @boyunJang in #2213
• 🌐 [i18n-KO] Translated guides/repository.md to Korean by @cjfghk5697 in #2124
• 🌐 [i18n-KO] Translated guides/model_cards.md to Korean" by @SeungAhSon in #2128
• 🌐 [i18n-KO] Translated guides/community.md to Korean by @seoulsky-field in #2126
• 🌐 [i18n-KO] Translated guides/cli.md to Korean by @harheem in #2131
• 🌐 [i18n-KO] Translated guides/search.md to Korean by @seoyoung-3060 in #2134
• 🌐 [i18n-KO] Translated guides/inference.md to Korean by @boyunJang in #2130
• 🌐 [i18n-KO] Translated guides/manage-spaces.md to Korean by @boyunJang in #2220
• 🌐 [i18n-KO] Translating guides/hf_file_system.md to Korean by @heuristicwave in #2146
• 🌐 [i18n-KO] Translated package_reference/hf_api.md to Korean by @fabxoe in #2165
• 🌐 [i18n-KO] Translated package_reference/mixins.md to Korean by @fabxoe in #2166
• 🌐 [i18n-KO] Translated guides/inference_endpoints.md to Korean by @usr-bin-ksh in #2164
• 🌐 [i18n-KO] Translated package_reference/utilities.md to Korean by @cjfghk5697 in #2196
• fix ko docs by @Wauplin (direct commit on main)
• 🌐 [i18n-KO] Translated package_reference/serialization.md to Korean by @seoyoung-3060 in #2233
• 🌐 [i18n-KO] Translated package_reference/hf_file_system.md to Korean by @SeungAhSon in #2174

🛠️ Misc improvements

User API

@bilgehanertan added support for 2 new routes:

• get_user_overview to retrieve high-level information about a user: username, avatar, number of models/datasets/Spaces, number of likes and upvotes, number of interactions in discussion, etc.

• User API endpoints by @bilgehanertan in #2147

CLI tag

@bilgehanertan added a new command to the CLI to handle tags. It is now possible to:

• tag a repo

>>> huggingface-cli tag Wauplin/my-cool-model v1.0
You are about to create tag v1.0 on model Wauplin/my-cool-model
Tag v1.0 created on Wauplin/my-cool-model

• retrieve the list of tags for a repo

>>> huggingface-cli tag Wauplin/gradio-space-ci -l --repo-type space
Tags for space Wauplin/gradio-space-ci:
0.2.2
0.2.1
0.2.0
0.1.2
0.0.2
0.0.1

• delete a tag on a repo

>>> huggingface-cli tag -d Wauplin/my-cool-model v1.0
You are about to delete tag v1.0 on model Wauplin/my-cool-model
Proceed? [Y/n] y
Tag v1.0 deleted on Wauplin/my-cool-model

For more details, check out the CLI guide.

• CLI Tag Functionality by @bilgehanertan in #2172

🧩 ModelHubMixin

This ModelHubMixin got a set of nice improvement to generate model cards and handle custom data types in the config.json file. More info in the integration guide.

• ModelHubMixin: more metadata + arbitrary config types + proper guide by @Wauplin in #2230
• Fix ModelHubMixin when class is a dataclass by @Wauplin in #2159
• Do not document private attributes of ModelHubMixin by @Wauplin in #2216
• Add support for pipeline_tag in ModelHubMixin by @Wauplin in #2228

⚙️ Other

In a shared environment, it is now possible to set a custom path HF_TOKEN_PATH as environment variable so that each user of the cluster has their own access token.

• Support HF_TOKEN_PATH as environment variable by @Wauplin in #2185

Thanks to @Y4suyuki and @lappemic, most custom errors defined in huggingface_hub are now aggregated in the same module. This makes it very easy...

[v0.22.2] Hot-fix: correctly handle proxies

Full Changelog: v0.22.1...v0.22.2

See #2155 and #2167 for more details.

[v0.22.1] Hot-fix: correctly handle dataclasses in ModelHubMixin

Fixed a bug breaking the SetFit integration.

What's Changed

• Fix use other chat completion providers by @Wauplin in #2153
• Fix ModelHubMixin when class is a dataclass by @Wauplin in #2159

Full Changelog: v0.22.0...v0.22.1

v0.22.0: Chat completion, inference types and hub mixins!

Discuss about the release in our Community Tab. Feedback is welcome!! 🤗

✨ InferenceClient

Support for inference tools continues to improve in huggingface_hub. At the menu in this release? A new chat_completion API and fully typed inputs/outputs!

Chat-completion API!

A long-awaited API has just landed in huggingface_hub! InferenceClient.chat_completion follows most of OpenAI's API, making it much easier to integrate with existing tools.

Technically speaking it uses the same backend as the text-generation task but requires a preprocessing step to format the list of messages into a single text prompt. The chat template is rendered server-side when models are powered by TGI, which is the case for most LLMs: Llama, Zephyr, Mistral, Gemma, etc. Otherwise, the templating happens client-side which requires minijinja package to be installed. We are actively working on bridging this gap, aiming at rendering all templates server-side in the future.

>>> from huggingface_hub import InferenceClient
>>> messages = [{"role": "user", "content": "What is the capital of France?"}]
>>> client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")

# Batch completion
>>> client.chat_completion(messages, max_tokens=100)
ChatCompletionOutput(
    choices=[
        ChatCompletionOutputChoice(
            finish_reason='eos_token',
            index=0,
            message=ChatCompletionOutputChoiceMessage(
                content='The capital of France is Paris. The official name of the city is "Ville de Paris" (City of Paris) and the name of the country\'s governing body, which is located in Paris, is "La République française" (The French Republic). \nI hope that helps! Let me know if you need any further information.'
            )
        )
    ],
    created=1710498360
)

# Stream new tokens one by one
>>> for token in client.chat_completion(messages, max_tokens=10, stream=True):
...     print(token)
ChatCompletionStreamOutput(choices=[ChatCompletionStreamOutputChoice(delta=ChatCompletionStreamOutputDelta(content='The', role='assistant'), index=0, finish_reason=None)], created=1710498504)
ChatCompletionStreamOutput(choices=[ChatCompletionStreamOutputChoice(delta=ChatCompletionStreamOutputDelta(content=' capital', role='assistant'), index=0, finish_reason=None)], created=1710498504)
(...)
ChatCompletionStreamOutput(choices=[ChatCompletionStreamOutputChoice(delta=ChatCompletionStreamOutputDelta(content=' may', role='assistant'), index=0, finish_reason=None)], created=1710498504)
ChatCompletionStreamOutput(choices=[ChatCompletionStreamOutputChoice(delta=ChatCompletionStreamOutputDelta(content=None, role=None), index=0, finish_reason='length')], created=1710498504)

• Implement InferenceClient.chat_completion + use new types for text-generation by @Wauplin in #2094
• Fix InferenceClient.text_generation for non-tgi models by @Wauplin in #2136
• #2153 by @Wauplin in #2153

Inference types

We are currently working towards more consistency in tasks definitions across the Hugging Face ecosystem. This is no easy job but a major milestone has recently been achieved! All inputs and outputs of the main ML tasks are now fully specified as JSONschema objects. This is the first brick needed to have consistent expectations when running inference across our stack: transformers (Python), transformers.js (Typescript), Inference API (Python), Inference Endpoints (Python), Text Generation Inference (Rust), Text Embeddings Inference (Rust), InferenceClient (Python), Inference.js (Typescript), etc.

Integrating those definitions will require more work but huggingface_hub is one of the first tools to integrate them. As a start, all InferenceClient return values are now typed dataclasses. Furthermore, typed dataclasses have been generated for all tasks' inputs and outputs. This means you can now integrate them in your own library to ensure consistency with the Hugging Face ecosystem. Specifications are open-source (see here) meaning anyone can access and contribute to them. Python's generated classes are documented here.

Here is a short example showcasing the new output types:

>>> from huggingface_hub import InferenceClient
>>> client = InferenceClient()
>>> client.object_detection("people.jpg"):
[
    ObjectDetectionOutputElement(
        score=0.9486683011054993,
        label='person',
        box=ObjectDetectionBoundingBox(xmin=59, ymin=39, xmax=420, ymax=510)
    ),
...
]

Note that those dataclasses are backward-compatible with the dict-based interface that was previously in use. In the example above, both ObjectDetectionBoundingBox(...).xmin and ObjectDetectionBoundingBox(...)["xmin"] are correct, even though the former should be the preferred solution from now on.

• Generate inference types + start using output types by @Wauplin in #2036
• Add = None at optional parameters by @LysandreJik in #2095
• Fix inference types shared between tasks by @Wauplin in #2125

🧩 ModelHubMixin

ModelHubMixin is an object that can be used as a parent class for the objects in your library in order to provide built-in serialization methods to upload and download pretrained models from the Hub. This mixin is adapted into a PyTorchHubMixin that can serialize and deserialize any Pytorch model. The 0.22 release brings its share of improvements to these classes:

1. Better support of init values. If you instantiate a model with some custom arguments, the values will be automatically stored in a config.json file and restored when reloading the model from pretrained weights. This should unlock integrations with external libraries in a much smoother way.
2. Library authors integrating the hub mixin can now define custom metadata for their library: library name, tags, document url and repo url. These are to be defined only once when integrating the library. Any model pushed to the Hub using the library will then be easily discoverable thanks to those tags.
3. A base modelcard is generated for each saved model. This modelcard includes default tags (e.g. model_hub_mixin) and custom tags from the library (see 2.). You can extend/modify this modelcard by overwriting the generate_model_card method.

>>> import torch
>>> import torch.nn as nn
>>> from huggingface_hub import PyTorchModelHubMixin


# Define your Pytorch model exactly the same way you are used to
>>> class MyModel(
...         nn.Module,
...         PyTorchModelHubMixin, # multiple inheritance
...         library_name="keras-nlp",
...         tags=["keras"],
...         repo_url="https://github.com/keras-team/keras-nlp",
...         docs_url="https://keras.io/keras_nlp/",
...         # ^ optional metadata to generate model card
...     ):
...     def __init__(self, hidden_size: int = 512, vocab_size: int = 30000, output_size: int = 4):
...         super().__init__()
...         self.param = nn.Parameter(torch.rand(hidden_size, vocab_size))
...         self.linear = nn.Linear(output_size, vocab_size)

...     def forward(self, x):
...         return self.linear(x + self.param)

# 1. Create model
>>> model = MyModel(hidden_size=128)

# Config is automatically created based on input + default values
>>> model._hub_mixin_config
{"hidden_size": 128, "vocab_size": 30000, "output_size": 4}

# 2. (optional) Save model to local directory
>>> model.save_pretrained("path/to/my-awesome-model")

# 3. Push model weights to the Hub
>>> model.push_to_hub("my-awesome-model")

# 4. Initialize model from the Hub => config has been preserved
>>> model = MyModel.from_pretrained("username/my-awesome-model")
>>> model._hub_mixin_config
{"hidden_size": 128, "vocab_size": 30000, "output_size": 4}

# Model card has been correctly populated
>>> from huggingface_hub import ModelCard
>>> card = ModelCard.load("username/my-awesome-model")
>>> card.data.tags
["keras", "pytorch_model_hub_mixin", "model_hub_mixin"]
>>> card.data.library_name
"keras-nlp"

For more details on how to integrate these classes, check out the integration guide.

• Fix ModelHubMixin: pass config when __init__ accepts **kwargs by @Wauplin in #2058
• [PyTorchModelHubMixin] Fix saving model with shared tensors by @NielsRogge in #2086
• Correctly inject config in PytorchModelHubMixin by @Wauplin in #2079
• Fix passing kwargs in PytorchHubMixin by @Wauplin in #2093
• Generate modelcard in ModelHubMixin by @Wauplin in #2080
• Fix ModelHubMixin: save config only if doesn't exist by @Wauplin in [#2105...

[v0.21.4] Hot-fix: Fix saving model with shared tensors

Release v0.21 introduced a breaking change make it impossible to save a PytorchModelHubMixin-based model that has shared tensors. This has been fixed in #2086.

Full Changelog: v0.21.3...v0.21.4

[v0.21.3] Hot-fix: ModelHubMixin pass config when `__init__` accepts `**kwargs`

More details in #2058.

Full Changelog: v0.21.2...v0.21.3

v0.21.2: hot-fix: [HfFileSystem] Fix glob with pattern without wildcards

See #2056. (+#2050 shipped as v0.21.1).

Full Changelog: v0.21.0...v0.21.2

v0.21.0: dataclasses everywhere, file-system, PyTorchModelHubMixin, serialization and more.

Discuss about the release in our Community Tab. Feedback welcome!! 🤗

🖇️ Dataclasses everywhere!

All objects returned by the HfApi client are now dataclasses!

In the past, objects were either dataclasses, typed dictionaries, non-typed dictionaries and even basic classes. This is now all harmonized with the goal of improving developer experience.

Kudos goes to the community for the implementation and testing of all the harmonization process. Thanks again for the contributions!

• Use dataclasses for all objects returned by HfApi #1911 by @Ahmedniz1 in #1974
• Updating HfApi objects to use dataclass by @Ahmedniz1 in #1988
• Dataclasses for objects returned hf api by @NouamaneELGueddarii in #1993

💾 FileSystem

The HfFileSystem class implements the fsspec interface to allow loading and writing files with a filesystem-like interface. The interface is highly used by the datasets library and this release will improve further the efficiency and robustness of the integration.

• Pass revision in path to AbstractBufferedFile init by @albertvillanova in #1948
• [HfFileSystem] Fix rm on branch by @lhoestq in #1957
• Retry fetching data on 502 error in HfFileSystem by @mariosasko in #1981
• Add HfFileSystemStreamFile by @lhoestq in #1967
• [HfFileSystem] Copy non lfs files by @lhoestq in #1996
• Add HfFileSystem.url method by @mariosasko in #2027

🧩 Pytorch Hub Mixin

The PyTorchModelHubMixin class let's you upload ANY pytorch model to the Hub in a few lines of code. More precisely, it is a class that can be inherited in any nn.Module class to add the from_pretrained, save_pretrained and push_to_hub helpers to your class. It handles serialization and deserialization of weights and configs for you and enables download counts on the Hub.

With this release, we've fixed 2 pain points holding back users from using this lib:

1. Configs are now better handled. The mixin automatically detects if the base class defines a config, saves it on the Hub and then injects it at load time, either as a dictionary or a dataclass depending on the base class's expectations.
2. Weights are now saved as .safetensors files instead of pytorch pickles for safety reasons. Loading from previous pytorch pickles is still supported but we are moving toward completely deprecating them (in a mid to long term plan).

• Better config support in ModelHubMixin by @Wauplin in #2001
• Use safetensors by default for PyTorchModelHubMixin by @bmuskalla in #2033

✨ InferenceClient improvements

Audio-to-audio task is now supported by both by the InferenceClient!

>>> from huggingface_hub import InferenceClient
>>> client = InferenceClient()
>>> audio_output = client.audio_to_audio("audio.flac")
>>> for i, item in enumerate(audio_output):
>>>     with open(f"output_{i}.flac", "wb") as f:
            f.write(item["blob"])

• Added audio to audio in inference client by @Ahmedniz1 in #2020

Also fixed a few things:

• Fix intolerance for new field in TGI stream response: 'index' by @danielpcox in #2006
• Fix optional model in tabular tasks by @Wauplin in #2018
• Added best_of to non-TGI ignored parameters by @dopc in #1949

📤 Model serialization

With the aim of harmonizing repo structures and file serialization on the Hub, we added a new module serialization with a first helper split_state_dict_into_shards that takes a state dict and split it into shards. Code implementation is mostly taken from transformers and aims to be reused by other libraries in the ecosystem. It seamlessly supports torch, tensorflow and numpy weights, and can be easily extended to other frameworks.

This is a first step in the harmonization process and more loading/saving helpers will be added soon.

• Framework-agnostic split_state_dict_into_shards helper by @Wauplin in #1938

📚 Documentation

🌐 Translations

Community is actively getting the job done to translate the huggingface_hub to other languages. We now have docs available in Simplified Chinese (here) and in French (here) to help democratize good machine learning!

• [i18n-CN] Translated some files to simplified Chinese #1915 by @2404589803 in #1916
• Update .github workflow to build cn docs on PRs by @Wauplin in #1931
• [i18n-FR] Translated files in french and reviewed them by @JibrilEl in #2024

Docs misc

• Document base_model in modelcard metadata by @Wauplin in #1936
• Update the documentation of add_collection_item by @FremyCompany in #1958
• Docs[i18n-en]: added pkgx as an installation method to the docs by @michaelessiet in #1955
• Added hf_transfer extra into setup.py and docs/ by @jamesbraza in #1970
• Documenting CLI default for download --repo-type by @jamesbraza in #1986
• Update repository.md by @xmichaelmason in #2010

Docs fixes

• Fix URL in get_safetensors_metadata docstring by @Wauplin in #1951
• Fix grammar by @Anthonyg5005 in #2003
• Fix doc by @jordane95 in #2013
• typo fix by @Decryptu in #2035

🛠️ Misc improvements

Creating a commit with an invalid README will fail early instead of uploading all LFS files before failing to commit.

• Fail early on invalid metadata by @Wauplin in #1934

Added a revision_exists helper, working similarly to repo_exists and file_exists:

>>> from huggingface_hub import revision_exists
>>> revision_exists("google/gemma-7b", "float16")
True
>>> revision_exists("google/gemma-7b", "not-a-revision")
False

• Add revision_exists helper by @Wauplin in #2042

InferenceClient.wait(...) now raises an error if the endpoint is in a failed state.

• raise on failed inference endpoint by @Wauplin in #1935

Improved progress bar when downloading a file

• improve http_get by @Wauplin in #1954

Other stuff:

• added will not echo message to the login token message by @vtrenton in #1925
• Raise if repo is disabled by @Wauplin in #1965
• Fix timezone in datetime parsing by @Wauplin in #1982
• retry on any 5xx on upload by @Wauplin in #2026

💔 Breaking changes

• Classes ModelFilter and DatasetFilter are deprecated when listing models and datasets in favor of a simpler API that lets you pass the parameters directly to list_models and list_datasets.

>>> from huggingface_hub import list_models, ModelFilter

# use
>>> list_models(language="zh")
# instead of 
>>> list_models(filter=ModelFilter(language="zh"))

Cleaner, right? ModelFilter and DatasetFilter will still be supported until v0.24 release.

• Deprecate ModelFilter/DatasetFilter by @druvdub in #2028
• List models tweaks by @julien-c in #2044

• In the inference client, ModelStatus.compute_type is not a string anymore but a dictionary with more detailed info...

0.20.3 hot-fix: Fix HfFolder login when env variable not set

This patch release fixes an issue when retrieving the locally saved token using huggingface_hub.HfFolder.get_token. For the record, this is a "planned to be deprecated" method, in favor of huggingface_hub.get_token which is more robust and versatile. The issue came from a breaking change introduced in #1895 meaning only 0.20.x is affected.

For more details, please refer to #1966.

Full Changelog: v0.20.2...v0.20.3

0.20.2 hot-fix: Fix concurrency issues in google colab login

A concurrency issue when using userdata.get to retrieve HF_TOKEN token led to deadlocks when downloading files in parallel. This hot-fix release fixes this issue by using a global lock before trying to get the token from the secrets vault. More details in #1953.

Full Changelog: v0.20.1...v0.20.2