๐Ÿ“ข [v0.24.0]: Inference, serialization and optimizations!

#7
by Wauplin HF staff - opened
Owner
โ€ข
edited Jul 18

Note: pre-release 0.24.0.rc0 is available on PyPI. Official release will occur in the coming days.

EDIT: 0.24.0 is out !

โšก๏ธ OpenAI-compatible inference client!

The InferenceClient's chat completion API is now fully compliant with OpenAI client. This means it's a drop-in replacement in your script:

- from openai import OpenAI
+ from huggingface_hub import InferenceClient

- client = OpenAI(
+ client = InferenceClient(
    base_url=...,
    api_key=...,
)


output = client.chat.completions.create(
    model="meta-llama/Meta-Llama-3-8B-Instruct",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Count to 10"},
    ],
    stream=True,
    max_tokens=1024,
)

for chunk in output:
    print(chunk.choices[0].delta.content)

Why switching to InferenceClient if you already use OpenAI then? Because it's better integrated with HF services, such as the Serverless Inference API and Dedicated Endpoints. Check out the more detailed answer in this HF Post.

For more details about OpenAI compatibility, check out this guide's section.

  • True OpenAI drop-in replacement by InferenceClient by @Wauplin in #2384
  • Promote chat_completion in inference guide by @Wauplin in #2366

(other) InferenceClient improvements

Some new parameters have been added to the InferenceClient, following the latest changes in our Inference API:

  • prompt_name, truncate and normalize in feature_extraction
  • model_id and response_format, in chat_completion
  • adapter_id in text_generation
  • hypothesis_template and multi_labels in zero_shot_classification

Of course, all of those changes are also available in the AsyncInferenceClient async equivalent ๐Ÿค—

  • Support truncate and normalize in InferenceClient by @Wauplin in #2270
  • Add prompt_name to feature-extraction + update types by @Wauplin in #2363
  • Send model_id in ChatCompletion request by @Wauplin in #2302
  • improve client.zero_shot_classification() by @MoritzLaurer in #2340
  • [InferenceClient] Add support for adapter_id (text-generation) and response_format (chat-completion) by @Wauplin in #2383

Added helpers for TGI servers:

  • get_endpoint_info to get information about an endpoint (running model, framework, etc.). Only available on TGI/TEI-powered models.
  • health_check to check health status of the server. Only available on TGI/TEI-powered models and only for InferenceEndpoint or local deployment. For serverless InferenceAPI, it's better to use get_model_status.

Other fixes:

  • image_to_text output type has been fixed
  • use wait-for-model to avoid been rate limited while model is not loaded
  • add proxies support
  • Fix InferenceClient.image_to_text output value by @Wauplin in #2285
  • Fix always None in text_generation output by @Wauplin in #2316
  • Add wait-for-model header when sending request to Inference API by @Wauplin in #2318
  • Add proxy support on async client by @noech373 in #2350
  • Remove jinja tips + fix typo in chat completion docstring by @Wauplin in #2368

๐Ÿ’พ Serialization

The serialization module introduced in v0.22.x has been improved to become the preferred way to serialize a torch model to disk. It handles how of the box sharding and safe serialization (using safetensors) with subtleties to work with shared layers. This logic was previously scattered in libraries like transformers, diffusers, accelerate and safetensors. The goal of centralizing it in huggingface_hub is to allow any external library to safely benefit from the same naming convention, making it easier to manage for end users.

>>> from huggingface_hub import save_torch_model
>>> model = ... # A PyTorch model

# Save state dict to "path/to/folder". The model will be split into shards of 5GB each and saved as safetensors.
>>> save_torch_model(model, "path/to/folder")

# Or save the state dict manually
>>> from huggingface_hub import save_torch_state_dict
>>> save_torch_state_dict(model.state_dict(), "path/to/folder") 

More details in the serialization package reference.

  • Serialization: support saving torch state dict to disk by @Wauplin in #2314
  • Handle shared layers in save_torch_state_dict + add save_torch_model by @Wauplin in #2373

Some helpers related to serialization have been made public for reuse in external libraries:

  • get_torch_storage_id
  • get_torch_storage_size
  • Support max_shard_size as string in split_state_dict_into_shards_factory by @SunMarc in #2286
  • Make get_torch_storage_id public by @Wauplin in #2304

๐Ÿ“ HfFileSystem

The HfFileSystem has been improved to optimize calls, especially when listing files from a repo. This is especially useful for large datasets like HuggingFaceFW/fineweb for faster processing and reducing risk of being rate limited.

  • [HfFileSystem] Less /paths-info calls by @lhoestq in #2271
  • Update token type definition and arg description in hf_file_system.py by @lappemic in #2278
  • [HfFileSystem] Faster fs.walk() by @lhoestq in #2346

Thanks to @lappemic, HfFileSystem methods are now properly documented. Check it out here!

  • Document more HfFilesyStem Methods by @lappemic in #2380

โœจ HfApi & CLI improvements

Commit API

A new mechanism has been introduced to prevent empty commits if no changes have been detected. Enabled by default in upload_file, upload_folder, create_commit and the huggingface-cli upload command. There is no way to force an empty commit.

  • Prevent empty commits if files did not change by @Wauplin in #2389

Resource groups

Resource Groups allow organizations administrators to group related repositories together, and manage access to those repos. It is now possible to specify a resource group ID when creating a repo:

from huggingface_hub import create_repo

create_repo("my-secret-repo", private=True, resource_group_id="66670e5163145ca562cb1988")

Webhooks API

Webhooks allow you to listen for new changes on specific repos or to all repos belonging to particular set of users/organizations (not just your repos, but any repo). With the Webhooks API you can create, enable, disable, delete, update, and list webhooks from a script!

from huggingface_hub import create_webhook

# Example: Creating a webhook
webhook = create_webhook(
    url="https://webhook.site/your-custom-url",
    watched=[{"type": "user", "name": "your-username"}, {"type": "org", "name": "your-org-name"}],
    domains=["repo", "discussion"],
    secret="your-secret"
)
  • [wip] Implement webhooks API by @lappemic in #2209

Search API

The search API has been slightly improved. It is now possible to:

  • filter datasets by tags
  • filter which attributes should be returned in model_info/list_models (and similarly for datasets/Spaces). For example, you can ask the server to return downloadsAllTime for all models.
>>> from huggingface_hub import list_models

>>> for model in list_models(library="transformers", expand="downloadsAllTime", sort="downloads", limit=5):
...     print(model.id, model.downloads_all_time)
MIT/ast-finetuned-audioset-10-10-0.4593 1676502301
sentence-transformers/all-MiniLM-L12-v2 115588145
sentence-transformers/all-MiniLM-L6-v2 250790748
google-bert/bert-base-uncased 1476913254
openai/clip-vit-large-patch14 590557280
  • Support filtering datasets by tags by @Wauplin in #2266
  • Support expand parameter in xxx_info and list_xxxs (model/dataset/Space) by @Wauplin in #2333
  • Add InferenceStatus to ExpandModelProperty_T by @Wauplin in #2388
  • Do not mention gitalyUid in expand parameter by @Wauplin in #2395

CLI

It is now possible to delete files from a repo using the command line:

Delete a folder:

>>> huggingface-cli repo-files Wauplin/my-cool-model delete folder/  
Files correctly deleted from repo. Commit: https://huggingface.co/Wauplin/my-cool-mo...

Use Unix-style wildcards to delete sets of files:

>>> huggingface-cli repo-files Wauplin/my-cool-model delete *.txt folder/*.bin 
Files correctly deleted from repo. Commit: https://huggingface.co/Wauplin/my-cool-mo...

ModelHubMixin

The ModelHubMixin, allowing for quick integration of external libraries with the Hub have been updated to fix some existing bugs and ease its use. Learn how to integrate your library from this guide.

  • Don't override 'config' in model_kwargs by @alexander-soare in #2274
  • Support custom kwargs for model card in save_pretrained by @qubvel in #2310
  • ModelHubMixin: Fix attributes lost in inheritance by @Wauplin in #2305
  • Fix ModelHubMixin coders by @gorold in #2291
  • Hot-fix: do not share tags between ModelHubMixin siblings by @Wauplin in #2394
  • Fix: correctly encode/decode config in ModelHubMixin if custom coders by @Wauplin in #2337

๐ŸŒ ๐Ÿ“š Documentation

Efforts from the Korean-speaking community continued to translate guides and package references to KO! Check out the result here.

  • ๐ŸŒ [i18n-KO] Translated package_reference/cards.md to Korean by @usr-bin-ksh in #2204
  • ๐ŸŒ [i18n-KO] Translated package_reference/community.md to Korean by @seoulsky-field in #2183
  • ๐ŸŒ [i18n-KO] Translated guides/collections.md to Korean by @usr-bin-ksh in #2192
  • ๐ŸŒ [i18n-KO] Translated guides/integrations.md to Korean by @cjfghk5697 in #2256
  • ๐ŸŒ [i18n-KO] Translated package_reference/environment_variables.md to Korean by @jungnerd in #2311
  • ๐ŸŒ [i18n-KO] Translated package_reference/webhooks_server.md to Korean by @fabxoe in #2344
  • ๐ŸŒ [i18n-KO] Translated guides/manage-cache.md to Korean by @cjfghk5697 in #2347

French documentation is also being updated, thanks to @JibrilEl!

  • [i18n-FR] Translated "Integrations" to french (sub PR of 1900) by @JibrilEl in #2329

A very nice illustration has been made by @severo to explain how hf:// urls works with the HfFileSystem object. Check it out here!

๐Ÿ’” Breaking changes

A few breaking changes have been introduced:

  • ModelFilter and DatasetFilter are completely removed. You can now pass arguments directly to list_models and list_datasets. This removes one level of complexity for the same result.
  • remove organization and name from update_repo_visibility. Please use a proper repo_id instead. This makes the method consistent with all other methods from HfApi.

These breaking changes have been announced with a regular deprecation cycle.

The legacy_cache_layout parameter (in hf_hub_download/snapshot_download) as well as cached_download, filename_to_url and url_to_filename helpers are now deprecated and will be removed in huggingface_hub==0.26.x. The proper way to download files is to use the current cache system with hf_hub_download/snapshot_download that have been in place for 2 years already.

  • Deprecate legacy_cache_layout parameter in hf_hub_download by @Wauplin in #2317

Small fixes and maintenance

โš™๏ธ fixes

  • Add comment to _send_telemetry_in_thread explaining it should not be removed by @freddyaboulton in #2264
  • feat: endpoints rename instances doc by @co42 in #2282
  • Fix FileNotFoundError in gitignore creation by @Wauplin in #2288
  • Close aiohttp client on error by @Wauplin in #2294
  • fix create_inference_endpoint by @nbroad1881 in #2292
  • Support custom_image in update_inference_endpoint by @Wauplin in #2306
  • Fix Repository if whoami call doesn't return an email by @Wauplin in #2320
  • print actual error message when failing to load a submodule by @kallewoof in #2342
  • Do not raise on .resume() if Inference Endpoint is already running by @Wauplin in #2335
  • Fix permission issue when downloading on root dir by @Wauplin in #2367
  • docs: update port for local doc preview in docs/README.md by @lappemic in #2382
  • Fix token=False not respected in file download by @Wauplin in #2386
  • Use extended path on Windows when downloading to local dir by @mlinke-ai in #2378
  • Add a default timeout for filelock by @edevil in #2391
  • Fix list_accepted_access_requests if grant user manually by @Wauplin in #2392
  • fix: Handle single return value. by 28Smiles in #2396

โš™๏ธ internal

  • Fix test: rename to open-llm-leaderboard + some cleaning by @Wauplin in #2295
  • Fix windows tests (git security update) by @Wauplin in #2296
  • Print correct webhook url when running in Spaces by @Wauplin in #2298
  • changed from --local_dir to --local-dir by @rao-pathangi in #2303
  • Update download badges in README by @Wauplin in #2309
  • Fix progress bar not always closed in file_download.py by @Wauplin in #2308
  • Make raises sections consistent in docstrings by @Wauplin in #2313
  • feat(ci): add trufflehog secrets detection by @McPatate in #2321
  • fix(ci): remove unnecessary permissions by @McPatate in #2322
  • Update _errors.py by @qgallouedec in #2354
  • Update ruff in CI by @Wauplin in #2365
  • Removing shebangs from files which are not supposed to be executable by @jpodivin in #2345
  • safetensors[torch] by @qgallouedec in #2371

Significant community contributions

The following contributors have made significant changes to the library over the last release:

  • @usr-bin-ksh
    • ๐ŸŒ [i18n-KO] Translated package_reference/cards.md to Korean (#2204)
    • ๐ŸŒ [i18n-KO] Translated guides/collections.md to Korean (#2192)
  • @seoulsky-field
    • ๐ŸŒ [i18n-KO] Translated package_reference/community.md to Korean (#2183)
  • @lappemic
    • Update token type definition and arg description in hf_file_system.py (#2278)
    • [wip] Implement webhooks API (#2209)
    • docs: update port for local doc preview in docs/README.md (#2382)
    • Document more HfFilesyStem Methods (#2380)
  • @rao-pathangi
    • changed from --local_dir to --local-dir (#2303)
  • @OlivierKessler01
    • fix/issue 2090 : Add a repo_files command, with recursive deletion. (#2280)
  • @qubvel
    • Support custom kwargs for model card in save_pretrained (#2310)
  • @gorold
    • Fix ModelHubMixin coders (#2291)
  • @cjfghk5697
    • ๐ŸŒ [i18n-KO] Translated guides/integrations.md to Korean (#2256)
    • ๐ŸŒ [i18n-KO] Translated guides/manage-cache.md to Korean (#2347)
  • @kallewoof
    • print actual error message when failing to load a submodule (#2342)
  • @jungnerd
    • ๐ŸŒ [i18n-KO] Translated package_reference/environment_variables.md to Korean (#2311)
  • @fabxoe
    • ๐ŸŒ [i18n-KO] Translated package_reference/webhooks_server.md to Korean (#2344)
  • @JibrilEl
    • [i18n-FR] Translated "Integrations" to french (sub PR of 1900) (#2329)
  • @noech373
    • Add proxy support on async client (#2350)
  • @jpodivin
    • Removing shebangs from files which are not supposed to be executable (#2345)
  • @mlinke-ai
    • Use extended path on Windows when downloading to local dir (#2378)
  • @edevil
    • Add a default timeout for filelock (#2391)

will there be any updates to the agent and tools/function calling ?

perhaps a wrapper class for local models ? so we can implement these fucntionitly locally , with ollama and ( automodels ) or Pipelines ? so a pipleine response can be simular to a open ai type response , indicatting stopping reasons ie ) tool call / Function call ) as well as inmplementing schema and json outputs for causal lm ( ie a wrapper) ....
enabling for the generation of automodels ( piplines) (LMS) to also use agent wrappers and hugging face (chain?)

this expansion or sset of plugin should be able to be used on inference client and transformers ( piplines ?) ( function calling / tools / schemas /json ) << as well as the problem of converting GGUF model back in to PRetrained tensors ( this is still having issues and does not work despite exisiting ?)

but other wise good work ?

Also maybe there is still a problem with bits and bytes compatablitys ?
perhaps this shoul be pushed to external refferencing ? as it is still a problem on windows ! ( and now the triton has been able to be installed on windows this bit is the broken bit !)

Owner

See my answer here about tools and agents. At the moment I don't think transformers's pipelines are returning the same return types as Text Generation Inference when ran locally. If you have a specific feature request, https://github.com/huggingface/transformers/issues is the way to go :)

For context, this huggingface_hub release (and the related discussion here) is purely related to the Python client to interact with the Hub. It does not contain the actual ML logic.

Sign up or log in to comment