Skip to content

Inference 2.0

Latest
Compare
Choose a tag to compare
@coyotte508 coyotte508 released this 19 Apr 11:35

The main features for this release are:

  • Support for Inference endpoints! credits to @vvmnnnkv
  • Custom requests via request and streamingRequest
  • Possibility to import the methods directly without the need to instantiate an HfInference class: great for tree-shaking
  • New NLP task: featureExtraction (the existing featureExtraction task was renamed to sentenceSimilarity, oops!), credits @radames

The other changes for recent versions are detailed at the end (including textGenerationStream for streaming text generation, ...)

Support for Inference Endpoints

Inference endpoints are the next step for using Inference API for a specific model in production.

The different tiers for inference are:

  • Inference API (no token): restrictive rate limits
  • Inference API - free account: usable rate limits
  • Inference API - PRO account: better rate limits
  • Inference Endpoints: Unlimited API calls, possibility to deploy on the cloud provider / VPC / infra of your choice, scaling

Here's how you can call an inference endpoint:

const inference = new HfInference("hf_...");

const gpt2 = inference.endpoint('https://xyz.eu-west-1.aws.endpoints.huggingface.cloud/gpt2');
const { generated_text } = await gpt2.textGeneration({inputs: 'The answer to the universe is'});

You can even use the free inference API backend with this syntax:

const endpoint = inference.endpoint("https://api-inference.huggingface.co/models/google/flan-t5-xxl");
const { generated_text } = await endpoint.textGeneration({
  inputs: "one plus two equals",
});

It's easy to switch between Inference API & Inference Endpoints. So easy, that you can even do this:

await inference.textGeneration({
  model: 'https://xyz.eu-west-1.aws.endpoints.huggingface.cloud/gpt2',
  inputs: 'The answer to the universe is'
});

Custom requests

@huggingface/inference supports tasks from https://huggingface.co/tasks, and is typed accordingly. But what if your model has additional inputs, or even custom inputs or outputs?

You can now use .request and .streamingRequest!

const output = await inference.request({
  inputs: "blablabla",
  parameters: {
    custom_parameter_1: ...,
    ...
  }
});

For streaming responses, use streamingRequest rather than request.

All existing tasks can use request or streamingRequest instead 🀯

const {generated_text} = await inference.textGeneration({model: "gpt2", inputs: "The answer to the universe is"});
// small output change for .textGeneration to .request: the raw response is actually an array
const [{generated_text}] = await inference.request({model: "gpt2", inputs: "The answer to the universe is"});

for await (const output of inference.textGenerationStream({
  model: "google/flan-t5-xxl", 
  inputs: "Repeat 'one two three four'"
})) {}
// is equivalent to
for await (const output of inference.streamingRequest({
  model: "google/flan-t5-xxl", 
  inputs: "Repeat 'one two three four'"
})) {}

Of course, request and streamingRequest can also be used with Inference Endpoints! Actually, if you make your own custom models and inputs / outputs for your business use case, it'll probably be what you use.

Individual imports & tree-shakability

You don't like the current API, you don't like classes, and want the strict minimum in your bundle? No need to say more, I know which frontend framework (or should I say library ;)) you use.

Don't worry, you can import individual functions - this release of @hugginface/inference is all about choice and flexibility:

import { textGeneration } from "@huggingface/inference";

await textGeneration({
  accessToken: "hf_...", // new param
  model: "gpt2", // or your own inference endpoint
  inputs: "The best, most efficient and purest frontend framework is: "
});

Breaking changes

  • questionAnswer and tableQuestionAnswer have been renamed to questionAnswering and tableQuestionAnswering
  • The existing featureExtraction has been renamed to sentenceSimilarity and a new featureExtraction was created πŸ™‡

Other changes from recent releases:

  • textGenerationStream to generate streaming content by returning an AsyncIterable. Yay for for await! Credits to @vvmnnnkv. Demo
  • imageToText to caption images among other things. Credits to @vvmnnnkv. Demo
  • Validation of outputs: Use request or streamingRequest to skip this validation. Credits to @mishig25