Uploading large models or datasets is challenging. We've already written some tips and tricks to facilitate the process but something was still missing. We are now glad to release the huggingface-cli upload-large-folder command. Consider it as a "please upload this no matter what, and be quick" command. Contrarily to huggingface-cli download, this new command is more opinionated and will split the upload into several commits. Multiple workers are started locally to hash, pre-upload and commit the files in a way that is resumable, resilient to connection errors, and optimized against rate limits. This feature has already been stress tested by the community over the last months to make it as easy and convenient to use as possible.
Here is how to use it:
huggingface-cli upload-large-folder <repo-id> <local-path> --repo-type=datasetEvery minute, a report is logged with the current status of the files and workers:
---------- 2024-04-26 16:24:25 (0:00:00) ----------
Files: hashed 104/104 (22.5G/22.5G) | pre-uploaded: 0/42 (0.0/22.5G) | committed: 58/104 (24.9M/22.5G) | ignored: 0
Workers: hashing: 0 | get upload mode: 0 | pre-uploading: 6 | committing: 0 | waiting: 0
---------------------------------------------------You can also run it from a script:
>>> from huggingface_hub import HfApi
>>> api = HfApi()
>>> api.upload_large_folder(
... repo_id="HuggingFaceM4/Docmatix",
... repo_type="dataset",
... folder_path="/path/to/local/docmatix",
... )
For more details about the command options, run:
huggingface-cli upload-large-folder --helpor visit the upload guide.
The search API have been updated. You can now list gated models and datasets, and filter models by their inference status (warm, cold, frozen).
More complete support for the expand[] parameter:
Organizations are now included when retrieving the user overview:
get_user_overview by @Wauplin in #2404get_user_followers and get_user_following are now paginated. This was not the case before, leading to issues for users with more than 1000 followers.
Added auth_check to easily verify if a user has access to a repo. It raises GatedRepoError if the repo is gated and the user don't have the permission or RepositoryNotFoundError if the repo does not exist or is private. If the method does not raise an error, you can assume the user has the permission to access the repo.
>>> from huggingface_hub import auth_check
>>> from huggingface_hub.utils import GatedRepoError, RepositoryNotFoundError
try:
auth_check("user/my-cool-model")
except GatedRepoError:
# Handle gated repository error
print("You do not have permission to access this gated repository.")
except RepositoryNotFoundError:
# Handle repository not found error
print("The repository was not found or you do not have access.")
auth_check by @cjfghk5697 in #2497It is now possible to set a repo as gated from a script:
>>> from huggingface_hub import HfApi
>>> api = HfApi()
>>> api.update_repo_settings(repo_id=repo_id, gated="auto") # Set to "auto", "manual" or False
update_repo_settings function to HfApi #2447 by @WizKnight in #2502A few improvements in the InferenceEndpoint API. It's now possible to set a scale_to_zero_timeout parameter + to configure secrets when creating or updating an Inference Endpoint.
The torch serialization module now supports tensor subclasses.
We also made sure that now the library is tested with both torch 1.x and 2.x to ensure compatibility.
weights_only=True by @Wauplin in #2488Breaking changes:
InferenceClient.conversational task has been removed in favor of InferenceClient.chat_completion. Also removed ConversationalOutput data class.InferenceClient output values are now dataclasses, not dictionaries.list_repo_likers is now paginated. This means the output is now an iterator instead of a list.Deprecation:
multi_commit: bool parameter in upload_folder is not deprecated, along the create_commits_on_pr. It is now recommended to use upload_large_folder instead. Thought its API and internals are different, the goal is still to be able to upload many files in several commits.Thanks to community feedback, we've been able to improve or fix significant things in both the InferenceClient and its async version AsyncInferenceClient. This fixes have been mainly focused on the OpenAI-compatible chat_completion method and the Inference Endpoints services.
stop parameter in text-generation instead of stop_sequences by @Wauplin in #2473InferenceClient for HF Nvidia NIM API by @Wauplin in #2482AsyncInferenceClient by @Wauplin in #2496When uploading a folder, we validate the README.md file before hashing all the files, not after. This should save some precious time when uploading large files and a corrupted model card.
Also, it is now possible to pass a --max-workers argument when uploading a folder from the CLI
All custom exceptions raised by huggingface_hub are now defined in huggingface_hub.errors module. This should make it easier to import them for your try/except statements.
At the same occasion, we've reworked how errors are formatted in hf_raise_for_status to print more relevant information to the users.
All constants in huggingface_hub are now imported as a module. This makes it easier to patch their values, for example in a test pipeline.
constants import to use module-level access #1172 by @WizKnight in #2453Other quality of life improvements:
HFSummaryWriter with 'hf-summary-writer' by @Wauplin in #2398RepoUrl info in CommitInfo object by @Wauplin in #2487touch() if OSError (to cache non existence of file) by @Wauplin in #2505The following contributors have made significant changes to the library over the last release:
auth_check (#2497)constants import to use module-level access #1172 (#2453)update_repo_settings function to HfApi #2447 (#2502)Full Changelog: https://github.com/huggingface/huggingface_hub/compare/v0.24.6...v0.24.7
For more details, refer to the related PR #2534.
Follow-up after https://github.com/huggingface/huggingface_hub/pull/2433 and v0.24.4 patch release. This release will definitely fix things.
Full Changelog: https://github.com/huggingface/huggingface_hub/compare/v0.24.4...v0.24.5
When downloading a file, the process was failing if the filesystem did not support either chmod or shutils.copy2 when moving a file from the tmp folder to the cache. This patch release fixes this. More details in https://github.com/huggingface/huggingface_hub/pull/2429.
Full Changelog: https://github.com/huggingface/huggingface_hub/compare/v0.24.3...v0.24.4
Fixing a bug in the chat completion URL to follow OpenAI standard https://github.com/huggingface/huggingface_hub/pull/2418. InferenceClient now works with urls ending with /, /v1 and /v1/chat/completions.
Full Changelog: https://github.com/huggingface/huggingface_hub/compare/v0.24.2...v0.24.3
See https://github.com/huggingface/huggingface_hub/pull/2413 for more details.
Creating an empty commit on a PR was failing due to a revision parameter been quoted twice. This patch release fixes it.
Full Changelog: https://github.com/huggingface/huggingface_hub/compare/v0.24.1...v0.24.2
This release fixes 2 things:
"[DONE]" message in chat stream (related to TGI update https://github.com/huggingface/text-generation-inference/pull/2221)See https://github.com/huggingface/huggingface_hub/pull/2410 for more details.
Full Changelog: https://github.com/huggingface/huggingface_hub/compare/v0.24.0...v0.24.1
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.
InferenceClient improvementsSome new parameters have been added to the InferenceClient, following the latest changes in our Inference API:
prompt_name, truncate and normalize in feature_extractionmodel_id and response_format, in chat_completionadapter_id in text_generationhypothesis_template and multi_labels in zero_shot_classificationOf course, all of those changes are also available in the AsyncInferenceClient async equivalent 🤗
prompt_name to feature-extraction + update types by @Wauplin in #2363adapter_id (text-generation) and response_format (chat-completion) by @Wauplin in #2383Added 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 fixedwait-for-model to avoid been rate limited while model is not loadedproxies supportThe 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.
save_torch_state_dict + add save_torch_model by @Wauplin in #2373Some helpers related to serialization have been made public for reuse in external libraries:
get_torch_storage_idget_torch_storage_sizemax_shard_size as string in split_state_dict_into_shards_factory by @SunMarc in #2286The 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.
hf_file_system.py by @lappemic in #2278fs.walk() by @lhoestq in #2346Thanks to @lappemic, HfFileSystem methods are now properly documented. Check it out here!
HfFilesyStem Methods by @lappemic in #2380A 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.
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")
resource_group_id in create_repo by @Wauplin in #2324Webhooks 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"
)
The search API has been slightly improved. It is now possible to:
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
expand parameter in xxx_info and list_xxxs (model/dataset/Space) by @Wauplin in #2333It 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...repo_files command, with recursive deletion. by @OlivierKessler01 in #2280The 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.
ModelHubMixin siblings by @Wauplin in #2394Efforts from the Korean-speaking community continued to translate guides and package references to KO! Check out the result here.
package_reference/cards.md to Korean by @usr-bin-ksh in #2204package_reference/community.md to Korean by @seoulsky-field in #2183guides/integrations.md to Korean by @cjfghk5697 in #2256package_reference/environment_variables.md to Korean by @jungnerd in #2311package_reference/webhooks_server.md to Korean by @fabxoe in #2344guides/manage-cache.md to Korean by @cjfghk5697 in #2347French documentation is also being updated, thanks to @JibrilEl!
A very nice illustration has been made by @severo to explain how hf:// urls works with the HfFileSystem object. Check it out here!
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.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.
legacy_cache_layout parameter in hf_hub_download by @Wauplin in #2317.resume() if Inference Endpoint is already running by @Wauplin in #2335docs/README.md by @lappemic in #2382safetensors[torch] by @qgallouedec in #2371The following contributors have made significant changes to the library over the last release:
package_reference/cards.md to Korean (#2204)package_reference/community.md to Korean (#2183)hf_file_system.py (#2278)docs/README.md (#2382)HfFilesyStem Methods (#2380)repo_files command, with recursive deletion. (#2280)guides/integrations.md to Korean (#2256)guides/manage-cache.md to Korean (#2347)package_reference/environment_variables.md to Korean (#2311)package_reference/webhooks_server.md to Korean (#2344)See https://github.com/huggingface/huggingface_hub/pull/2394 for more details.
Full Changelog: https://github.com/huggingface/huggingface_hub/compare/v0.23.4...v0.23.5
Includes:
Full Changelog: https://github.com/huggingface/huggingface_hub/compare/v0.23.3...v0.23.4
Release 0.23.0 introduced a breaking change in InferenceClient.text_generation. When details=True is passed, the details attribute in the output is always None. The patch release fixes this. See https://github.com/huggingface/huggingface_hub/pull/2316 for more details.
Full Changelog: https://github.com/huggingface/huggingface_hub/compare/v0.23.2...v0.23.3
split_state_dict_into_shards_factory now accepts string values as max_shard_size (ex: "5MB"), in addition to integer values. Related PR: https://github.com/huggingface/huggingface_hub/pull/2286.
Full Changelog: https://github.com/huggingface/huggingface_hub/compare/v0.23.1...v0.23.2
See https://github.com/huggingface/huggingface_hub/pull/2271 for more details.
Full Changelog: https://github.com/huggingface/huggingface_hub/compare/v0.23.0...v0.23.1
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:
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/phi3With 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).
.huggingface/ folder to .cache/huggingface/ by @Wauplin in #2262InferenceClientIt 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.
Mention more chat-completion task instead of conversation in documentation.
chat_completion and remove conversational from Inference guide by @Wauplin in #2215chat-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.
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
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.
guides/webhooks_server.md to Korean by @nuatmochoi in #2145reference/login.md to Korean by @SeungAhSon in #2151package_reference/tensorboard.md to Korean by @fabxoe in #2173package_reference/inference_client.md to Korean by @cjfghk5697 in #2178reference/inference_endpoints.md to Korean by @harheem in #2180package_reference/file_download.md to Korean by @seoyoung-3060 in #2184package_reference/cache.md to Korean by @nuatmochoi in #2191package_reference/collections.md to Korean by @boyunJang in #2214package_reference/inference_types.md to Korean by @fabxoe in #2171guides/upload.md to Korean by @junejae in #2139reference/repository.md to Korean by @junejae in #2189package_reference/space_runtime.md to Korean by @boyunJang in #2213guides/repository.md to Korean by @cjfghk5697 in #2124guides/model_cards.md to Korean" by @SeungAhSon in #2128guides/community.md to Korean by @seoulsky-field in #2126guides/cli.md to Korean by @harheem in #2131guides/search.md to Korean by @seoyoung-3060 in #2134guides/inference.md to Korean by @boyunJang in #2130guides/manage-spaces.md to Korean by @boyunJang in #2220guides/hf_file_system.md to Korean by @heuristicwave in #2146package_reference/hf_api.md to Korean by @fabxoe in #2165package_reference/mixins.md to Korean by @fabxoe in #2166guides/inference_endpoints.md to Korean by @usr-bin-ksh in #2164package_reference/utilities.md to Korean by @cjfghk5697 in #2196@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.@bilgehanertan added a new command to the CLI to handle tags. It is now possible to:
>>> 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
>>> 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
>>> 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.
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 #2230In 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.
HF_TOKEN_PATH as environment variable by @Wauplin in #2185Thanks to @Y4suyuki and @lappemic, most custom errors defined in huggingface_hub are now aggregated in the same module. This makes it very easy to import them from from huggingface_hub.errors import ....
Fixed HFSummaryWriter (class to seamlessly log tensorboard events to the Hub) to work with either tensorboardX or torch.utils implementation, depending on the user setup.
Speed to list files using HfFileSystem has been drastically improved, thanks to @awgr. The values returned from the cache are not deep-copied anymore, which was unfortunately the part taking the most time in the process. If users want to modify values returned by HfFileSystem, they would need to copy them before-hand. This is expected to be a very limited drawback.
Progress bars in huggingface_hub got some flexibility!
It is now possible to provide a name to a tqdm bar (similar to logging.getLogger) and to enable/disable only some progress bars. More details in this guide.
>>> from huggingface_hub.utils import tqdm, disable_progress_bars
>>> disable_progress_bars("peft.foo")
# No progress bars for `peft.boo.bar`
>>> for _ in tqdm(range(5), name="peft.foo.bar"):
... pass
# But for `peft` yes
>>> for _ in tqdm(range(5), name="peft"):
... pass
100%|█████████████████| 5/5 [00:00<00:00, 117817.53it/s]
--local-dir-use-symlink and --resume-downloadAs part of the download process revamp, some breaking changes have been introduced. However we believe that the benefits outweigh the change cost. Breaking changes include:
.cache/huggingface/ folder is not present at the root of the local dir. It only contains file locks, metadata and partially downloaded files. If you need to, you can safely delete this folder without corrupting the data inside the root folder. However, you should expect a longer recovery time if you try to re-run your download command.--local-dir-use-symlink is not in used anymore and will be ignored. It is not possible anymore to symlinks your local dir with the cache directory. Thanks to the .cache/huggingface/ folder, it shouldn't be needed anyway.--resume-download has been deprecated and will be ignored. Resuming failed downloads is now activated by default all the time. If you need to force a new download, use --force-download.As part of #2237 (Grammar and Tools support), we've updated the return value from InferenceClient.chat_completion and InferenceClient.text_generation to match exactly TGI output. The attributes of the returned objects did not change but the classes definition themselves yes. Expect errors if you've previously had from huggingface_hub import TextGenerationOutput in your code. This is however not the common usage since those objects are already instantiated by huggingface_hub directly.
Some other breaking changes were expected (and announced since 0.19.x):
list_files_info is definitively removed in favor of get_paths_info and list_repo_treeWebhookServer.run is definitively removed in favor of WebhookServer.launchapi_endpoint in ModelHubMixin push_to_hub's method is definitively removed in favor of the HF_ENDPOINT environment variableCheck #2156 for more details.
hf_file_system by @Wauplin in #2253updatedRefs in WebhookPayload by @Wauplin in #2169TestHfHubDownloadRelativePaths + implicit delete folder is ok by @Wauplin in #2259The following contributors have made significant changes to the library over the last release:
guides/community.md to Korean (#2126)guides/hf_file_system.md to Korean (#2146)guides/inference_endpoints.md to Korean (#2164)Fixed a bug breaking the SetFit integration.
Full Changelog: https://github.com/huggingface/huggingface_hub/compare/v0.22.0...v0.22.1
Discuss about the release in our Community Tab. Feedback is welcome!! 🤗
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!
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)
InferenceClient.chat_completion + use new types for text-generation by @Wauplin in #2094We 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.
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:
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.
ModelHubMixin: pass config when __init__ accepts **kwargs by @Wauplin in #2058PytorchModelHubMixin by @Wauplin in #2079ModelHubMixin by @Wauplin in #2080HfFileSystem download speed was limited by some internal logic in fsspec. We've now updated the get_file and read implementations to improve their download speed to a level similar to hf_hub_download.
We are aiming at moving all errors raised by huggingface_hub into a single module huggingface_hub.errors to ease the developer experience. This work has been started as a community contribution from @Y4suyuki.
HfApi class now accepts a headers parameters that is then passed to every HTTP call made to the Hub.
📚 More documentation in Korean!
package_reference/overview.md to Korean by @jungnerd in #2113The new types returned by InferenceClient methods should be backward compatible, especially to access values either as attributes (.my_field) or as items (i.e. ["my_field"]). However, dataclasses and dicts do not always behave exactly the same so might notice some breaking changes. Those breaking changes should be very limited.
ModelHubMixin internals changed quite a bit, breaking some use cases. We don't think those use cases were in use and changing them should really benefit 99% of integrations. If you witness any inconsistency or error in your integration, please let us know and we will do our best to mitigate the problem. One of the biggest change is that the config values are not attached to the mixin instance as instance.config anymore but as instance._model_hub_mixin. The .config attribute has been mistakenly introduced in 0.20.x so we hope it has not been used much yet.
huggingface_hub.file_download.http_user_agent has been removed in favor of the officially document huggingface_hub.utils.build_hf_headers. It was a deprecated method since 0.18.x.
The CI pipeline has been greatly improved, especially thanks to the efforts from @bmuskalla. Most tests are now passing in under 3 minutes, against 8 to 10 minutes previously. Some long-running tests have been greatly simplified and all tests are now ran in parallel with python-xdist, thanks to a complete decorrelation between them.
We are now also using the great uv installer instead of pip in our CI, which saves around 30-40s per pipeline.
python-xdist on all tests by @bmuskalla in #2059The following contributors have made significant changes to the library over the last release:
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 https://github.com/huggingface/huggingface_hub/pull/2086.
Full Changelog: https://github.com/huggingface/huggingface_hub/compare/v0.21.3...v0.21.4
More details in https://github.com/huggingface/huggingface_hub/pull/2058.
Full Changelog: https://github.com/huggingface/huggingface_hub/compare/v0.21.2...v0.21.3