What does HF_HOME actually control?

HF_HOME sets the root directory the huggingface_hub and transformers libraries use to cache downloaded models, tokenizers, and datasets. Pointing it at a project-local directory (instead of the default ~/.cache/huggingface) keeps large model downloads out of your home directory and makes cache cleanup per-project.

When should I enable trust_remote_code?

Only when the model repository ships custom Python modeling code that transformers needs to execute to run the model (common for some newer architectures), and only for authors you trust. That code runs on your machine with your permissions, so enabling it for an unfamiliar repo is a real code-execution risk, not just a compatibility flag.

What's the difference between the local and remote (Inference Endpoint) modes here?

Local mode downloads the model's weights and runs inference on your own machine via transformers' pipeline() function. Remote mode calls a Hugging Face Inference Endpoint you've already deployed, using huggingface_hub's InferenceClient, so the actual computation happens on Hugging Face's infrastructure instead of yours.

Does trust_remote_code apply to Inference Endpoints too?

No. It's a local transformers loading option; when calling a hosted Inference Endpoint, whatever code the endpoint runs was already configured at deployment time, so the flag has no effect on the client side and the remote snippet omits it.

Hugging Face Config Generator - Local or Inference Endpoint

Overview

Introduction

Getting a Hugging Face model running usually means deciding where model weights should be cached, whether the model needs custom remote code to load, and whether you're running it locally or calling a hosted endpoint, three small decisions that are easy to get wrong on a first try.

This tool generates the HF_HOME/HF_TOKEN environment setup plus a matching Python snippet for either a local transformers pipeline or a remote Inference Endpoint call.

What Is Hugging Face Config Generator?

A generator for Hugging Face environment configuration (HF_HOME cache directory, an HF_TOKEN reference) paired with a Python snippet, either a local `transformers.pipeline()` call with a trust_remote_code toggle, or a remote `huggingface_hub.InferenceClient` call for a deployed Inference Endpoint.

The local snippet includes a visible caution comment whenever trust_remote_code is enabled, since that setting executes code shipped in the model repository.

How Hugging Face Config Generator Works

You choose local or remote mode, set the cache directory, enter a model ID, and (in local mode) toggle trust_remote_code. The generator validates the model ID shape and required fields.

In local mode it emits an HF_HOME/HF_TOKEN block and a `pipeline()` snippet with the trust_remote_code value set accordingly. In remote mode it emits the same env block and an `InferenceClient` snippet targeting the model ID as a deployed endpoint, noting that trust_remote_code doesn't apply there.

When To Use Hugging Face Config Generator

Use it when starting a new script that loads a Hugging Face model locally, to get the cache directory and pipeline call right the first time.

Use remote mode when you've already deployed a model to a Hugging Face Inference Endpoint and need the client snippet to call it.

Often used alongside Ollama Docker Compose Generator, Vector Database Config Generator and OpenAI Environment Generator.

Features

Advantages

Keeps model cache location explicit and project-local via HF_HOME, instead of silently filling the user's home cache.
Surfaces a clear caution note whenever trust_remote_code is enabled, rather than leaving that risk undocumented in the code.
Covers both the local inference and hosted Inference Endpoint paths with the correct client for each.

Limitations

The local snippet defaults to the generic text-generation pipeline task; other tasks (e.g. image-classification, text-to-speech) need the task string changed manually.
Doesn't validate that the model ID actually exists on the Hugging Face Hub or that an Inference Endpoint is deployed for it.

Examples

Local text-generation pipeline without remote code

Input

mode: local, cacheDir: ./.cache/huggingface, modelId: mistralai/Mistral-7B-Instruct-v0.3, trustRemoteCode: false

Output

# .env
HF_HOME=./.cache/huggingface
HF_TOKEN=hf_your_token_here

import os
from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="mistralai/Mistral-7B-Instruct-v0.3",
    trust_remote_code=False,
    token=os.environ.get("HF_TOKEN"),
)

result = pipe("Hello, how are you?")
print(result)

No caution comment is added because trust_remote_code is false; the token is read from HF_TOKEN via os.environ.get so a missing token doesn't crash public-model loads.

Best Practices & Notes

Best Practices

Set HF_HOME to a project-local directory in larger teams so each project's model cache can be cleared or backed up independently.
Leave trust_remote_code off by default and only enable it per-model after checking the repository's code, not as a blanket setting.
Use the remote Inference Endpoint mode for production traffic once a model is validated, rather than serving local pipeline() inference from a production process.

Developer Notes

HF_HOME is read by both `huggingface_hub` and `transformers` to locate the shared cache for hub downloads (models, tokenizers, datasets); it supersedes the older, more granular TRANSFORMERS_CACHE/HF_DATASETS_CACHE variables in current library versions. The local snippet targets the `transformers.pipeline()` high-level API; the remote snippet targets `huggingface_hub.InferenceClient`, which calls a deployed Inference Endpoint's HTTP API under the hood.

Hugging Face Config Generator Use Cases

Bootstrapping a script that runs a Hugging Face model locally with a project-scoped cache directory
Generating a client snippet to call an already-deployed Hugging Face Inference Endpoint
Documenting the trust_remote_code decision and its risk for a specific model in a reproducible snippet

Common Mistakes

Enabling trust_remote_code for an unfamiliar model repository without reviewing its code first.
Leaving HF_HOME unset and being surprised when model downloads silently fill the default home-directory cache.

Tips

If you only need a quick local test, leave HF_TOKEN unset for public models, the pipeline() call still works since token is read with os.environ.get rather than a hard os.environ[...] lookup.

Hugging Face Config Generator