hf-mem

Estimate inference memory requirements for Hugging Face models.

Demo: taubyte.com/tools/hf-mem — try it in the browser with any model ID, optional token, and GPU selection.

Installation

npm install hf-mem

Usage

import { run } from 'hf-mem';

// Basic usage
const result = await run('MiniMaxAI/MiniMax-M2');

console.log(result.metadata.bytesCount); // Total bytes
console.log(result.metadata.paramCount); // Total parameters

// With options
const resultWithOptions = await run('MiniMaxAI/MiniMax-M2', 'main', {
  token: 'hf_...', // Optional: for private models
  jsonOutput: true // Optional: include JSON output
});

if (resultWithOptions.json) {
  console.log(resultWithOptions.json);
}

API

run(modelId, revision?, options?)

Estimates memory requirements for a Hugging Face model.

Parameters:

• modelId (string): The Hugging Face model ID (e.g., 'MiniMaxAI/MiniMax-M2')
• revision (string, optional): Model revision (default: "main")
• options (object, optional):
  • token (string, optional): Hugging Face token for private models
  • jsonOutput (boolean, optional): If true, includes JSON output in result

Returns:

• Promise<{ metadata: SafetensorsMetadata; json?: any }>

Example:

const result = await run('bert-base-uncased', 'main', {
  token: process.env.HF_TOKEN,
  jsonOutput: false
});

// Access metadata
console.log(`Total parameters: ${result.metadata.paramCount}`);
console.log(`Total bytes: ${result.metadata.bytesCount}`);
console.log(`Components:`, Object.keys(result.metadata.components));

Types

import type {
  SafetensorsMetadata,
  ComponentMetadata,
  DtypeMetadata,
  SafetensorsDtypes
} from 'hf-mem';

import { RuntimeError } from 'hf-mem';

Supported Model Types

• Transformers models: Standard PyTorch models with model.safetensors or sharded variants
• Diffusers models: Models with model_index.json and component-based structure
• Sentence Transformers: Models with additional Dense layers

How It Works

The library:

1. Fetches the model's file tree from the Hugging Face Hub API
2. Uses HTTP Range requests to fetch only the metadata portion of safetensors files (first ~100KB)
3. Parses the binary safetensors format to extract tensor metadata
4. Calculates memory requirements based on tensor shapes and data types

Browser and Node.js Support

This package works in both Node.js (18+) and modern browsers that support the fetch API.

License

MIT