3.4.0 version of transformer does not respect env.localPath

[djaffer]

System Info

The version 3.4.0 breaks the use local model. The version 3.3.3 works perfectly.

I have a path like

import {env, pipeline } from '@huggingface/transformer";

env.localModelPath = '../model-dir';

Environment/Platform

• Website/web-app
• Browser extension
• Server-side (e.g., Node.js, Deno, Bun)
• Desktop app (e.g., Electron)
• Other (e.g., VSCode extension)

Description

The version 3.4.0 breaks the use local model. The version 3.3.3 works perfectly.

I have a path like

import {env, pipeline } from '@huggingface/transformer";

env.localModelPath = '../model-dir';

Reproduction

The version 3.4.0 breaks the use local model. The version 3.3.3 works perfectly.

I have a path like

import {env, pipeline } from '@huggingface/transformer";

env.localModelPath = '../model-dir';

env.allowRemoteModels = false;
const model = await pipeline(
'feature-extraction',
'bge-base-en',
);

[xenova]

Hi there 👋

Are you running in Node.js or in-browser?

If Node.js (or other server-side environment), can you try use the relative path for the model name (e.g., ./model-dir/bge-base-en).

The change made in v3.4.0 was meant to better align with usage of the python library.

[djaffer]

I tried it same issue even with absolute path if it is in different directory it didn't work.

[xenova]

Are you running in Node.js or in-browser?

This will help debug :) Also, if you can share the error message you are seeing, that will be helpful.

[djaffer]

in node js.

throw Error(\local_files_only=true` or `env.allowRemoteModels=false` and file was not found locally at "${localPath}".`);
^

Error: local_files_only=true or env.allowRemoteModels=false and file was not found locally at "bge-base-en-v1.5/tokenizer.json".
at getModelFile (file:///project/node_modules/@huggingface/transformers/dist/transformers.node.mjs:30145:27)
at async getModelJSON (file:///project/node_modules/@huggingface/transformers/dist/transformers.node.mjs:30278:20)
at async Promise.all (index 0)
at async loadTokenizer (file:///project/node_modules/@huggingface/transformers/dist/transformers.node.mjs:23501:18) at async AutoTokenizer.from_pretrained (file:///project/node_modules/@huggingface/transformers/dist/transformers.node.mjs:27804:50)
at async Promise.all (index 0)
at async loadItems (file:///project/node_modules/@huggingface/transformers/dist/transformers.node.mjs:23372:5)
at async Module.pipeline (file:///project/node_modules/@huggingface/transformers/dist/transformers.node.mjs:23302:21)
at async HuggingFaceTransformersEmbeddings.runEmbedding (/project/node_modules/@langchain/community/dist/embeddings/huggingface_transformers.cjs:99:22)
at async HuggingFaceTransformersEmbeddings.embedQuery (/project/node_modules/@langchain/community/dist/embeddings/huggingface_transformers.cjs:93:22)

[vishnusureshperumbavoor]

Required Directory Structure

model-dir/
└── bge-base-en-v1.5/        # Note the version suffix
    ├── config.json
    ├── model.safetensors
    ├── tokenizer.json
    └── tokenizer_config.json

[djaffer]

yes I tried it. How is it possible that once I move to latest it stops working. Definitely it version issue.

model-dir/
└── bge-base-en-v1.5/ # Note the version suffix
├── config.json
├── model.safetensors
├── tokenizer.json
└── tokenizer_config.json

[catherine917]

I got the same issue, too

[nemphys]

Same here, it seems that the recent change in hub.js in the commit for the external data support messed up localModelPath; it now tries to load the local file without prepending the variable's path and therefore fails.
Maybe

let localPath = requestURL;

should be

let localPath = cachePath;

?

[xenova]

#1268 should fix this! Please let me know if it does; I'm planning a patch release for this shortly (3.4.2)