Feature Request: Expose an API to configure `@huggingface/transformers` `env` inside the Web Worke

[Wyatex]

Description

Currently, the @huggingface/transformers library is imported and executed entirely within a Web Worker. Because Web Workers have isolated execution environments, the exported env object from @huggingface/transformers is trapped inside the worker thread.

As a result, developers using this library from the main thread cannot access or modify the env variables (e.g., env.allowLocalModels, etc.). This strict isolation makes it difficult to customize the behavior of the transformers library dynamically.

Solution

Add support for an additional env parameter in the createOptions argument of createXxxProvider to be passed to the worker. Furthermore, after instantiating new Worker, automatically check if a custom env was provided, and automatically call the worker's env modification method.

[nekomeowww]

Oh... I never though it would have such limitations. Let me fix this. Would you mind share your setup of the xsai-transformers? Node.js environment? Or Web browser? And what env do you need?

[Wyatex]

Currently, I'm only using this in a browser environment. Below are the configurations I might use:

import { env } from "@huggingface/transformers";
env.allowLocalModels = true;
env.allowRemoteModels = false;
env.localModelPath = "/xxx";

and

env.remoteHost = 'http://127.0.0.1:8080'
env.remotePathTemplate = '{model}'

So, should we extend the parameters of createXxxProvider to be as follows:

const embedProvider = createEmbedProvider({
  baseURL: `xsai-transformers:///?worker-url=${embedWorkerURL}`,
  env: {
    remoteHost: 'http://127.0.0.1:8080',
    remotePathTemplate: '{model}',
  }
})

Or should we pass it when calling loadXxx ? Like this:

await embedProvider.loadEmbed(modelId, {
  // ...
  env: {
    remoteHost: 'http://127.0.0.1:8080',
    remotePathTemplate: '{model}',
  }
})

Sure. Considering others might need custom env setups, I think we should align with the official env configuration specs.

[Wyatex]

To use it on an internal network, I might also need the following configuration: env.backends.onnx.wasm.wasmPaths = '/wasm/'