🤗 Hugging Face Hub API

Official utilities to use the Hugging Face Hub API.

Install

1pnpm add @huggingface/hub
2
3npm add @huggingface/hub
4
5yarn add @huggingface/hub

Deno

1// esm.sh
2import { uploadFiles, listModels } from "https://esm.sh/@huggingface/hub"
3// or npm:
4import { uploadFiles, listModels } from "npm:@huggingface/hub"

Check out the full documentation.

Usage

For some of the calls, you need to create an account and generate an access token.

Learn how to find free models using the hub package in this interactive tutorial.

1import * as hub from "@huggingface/hub";
2import type { RepoDesignation } from "@huggingface/hub";
3
4const repo: RepoDesignation = { type: "model", name: "myname/some-model" };
5
6const {name: username} = await hub.whoAmI({accessToken: "hf_..."});
7
8for await (const model of hub.listModels({search: {owner: username}, accessToken: "hf_..."})) {
9  console.log("My model:", model);
10}
11
12const specificModel = await hub.modelInfo({name: "openai-community/gpt2"});
13await hub.checkRepoAccess({repo, accessToken: "hf_..."});
14
15await hub.createRepo({ repo, accessToken: "hf_...", license: "mit" });
16
17await hub.uploadFiles({
18  repo,
19  accessToken: "hf_...",
20  files: [
21    // path + blob content
22    {
23      path: "file.txt",
24      content: new Blob(["Hello World"]),
25    },
26    // Local file URL
27    pathToFileURL("./pytorch-model.bin"),
28    // Web URL
29    new URL("https://huggingface.co/xlm-roberta-base/resolve/main/tokenizer.json"),
30    // Path + Web URL
31    {
32      path: "myfile.bin",
33      content: new URL("https://huggingface.co/bert-base-uncased/resolve/main/pytorch_model.bin")
34    }
35    // Can also work with native File in browsers
36  ],
37});
38
39// or
40
41for await (const progressEvent of await hub.uploadFilesWithProgress({
42  repo,
43  accessToken: "hf_...",
44  files: [
45    ...
46  ],
47})) {
48  console.log(progressEvent);
49}
50
51await hub.deleteFile({repo, accessToken: "hf_...", path: "myfile.bin"});
52
53await (await hub.downloadFile({ repo, path: "README.md" })).text();
54
55for await (const fileInfo of hub.listFiles({repo})) {
56  console.log(fileInfo);
57}
58
59await hub.deleteRepo({ repo, accessToken: "hf_..." });

OAuth Login

It's possible to login using OAuth ("Sign in with HF").

This will allow you get an access token to use some of the API, depending on the scopes set inside the Space or the OAuth App.

1import { oauthLoginUrl, oauthHandleRedirectIfPresent } from "@huggingface/hub";
2
3const oauthResult = await oauthHandleRedirectIfPresent();
4
5if (!oauthResult) {
6  // If the user is not logged in, redirect to the login page
7  window.location.href = await oauthLoginUrl();
8}
9
10// You can use oauthResult.accessToken, oauthResult.accessTokenExpiresAt and oauthResult.userInfo
11console.log(oauthResult);

Checkout the demo: https://huggingface.co/spaces/huggingfacejs/client-side-oauth

Hugging face cache

The @huggingface/hub package provide basic capabilities to scan the cache directory. Learn more about Manage huggingface_hub cache-system.

scanCacheDir

You can get the list of cached repositories using the scanCacheDir function.

1import { scanCacheDir } from "@huggingface/hub";
2
3const result = await scanCacheDir();
4
5console.log(result);

Note: this does not work in the browser

downloadFileToCacheDir

You can cache a file of a repository using the downloadFileToCacheDir function.

1import { downloadFileToCacheDir } from "@huggingface/hub";
2
3const file = await downloadFileToCacheDir({
4  repo: 'foo/bar',
5  path: 'README.md'
6});
7
8console.log(file);

Note: this does not work in the browser

snapshotDownload

You can download an entire repository at a given revision in the cache directory using the snapshotDownload function.

1import { snapshotDownload } from "@huggingface/hub";
2
3const directory = await snapshotDownload({
4  repo: 'foo/bar',
5});
6
7console.log(directory);

The code use internally the downloadFileToCacheDir function.

Note: this does not work in the browser

Performance considerations

When uploading large files, you may want to run the commit calls inside a worker, to offload the sha256 computations.

Remote resources and local files should be passed as URL whenever it's possible so they can be lazy loaded in chunks to reduce RAM usage. Passing a File inside the browser's context is fine, because it natively behaves as a Blob.

Under the hood, @huggingface/hub uses a lazy blob implementation to load the file.

Dependencies

• @huggingface/tasks : Typings only