localgrid 0.1.1

A library for inspecting local LLM providers and tokenizers.

LocalGrid

A simple Python library to inspect and get metadata for local LLMs, like token limits and token counts.

I built this because I needed a way to get accurate info for local models without having to make any web calls. All the data and tokenizers are bundled directly into the package.

localgrid on PyPI

Supports Python >=3.8

Key Features

• Fully Local: No internet connection needed after installation.
• Accurate Token Counting: Uses the real tokenizer for a given model, not just a guess.
• Bundled Tokenizers: All the necessary tokenizer files are included in the package.
• Simple API: Just a few functions to get what you need.

Data Source

The model data (like context limits) was gathered by scraping and formatting information from Ollama and lm-studios public model library's. This data is saved in a JSON file (localgrid_cache.json) inside the package.

view the models

or just visit ollama and lm-studio

Installation

pip install localgrid

Quick Start & Usage

The library has three main functions you'll probably use.

1. get_context_limit

Gets the total context size (token limit) for a model.

from localgrid import get_context_limit

# Get the context limit for llama3.1:latest
limit = get_context_limit("llama3.1:latest")#ollama format

print(f"llama3.1:latest limit: {limit}")
# Output: llama3.1:latest limit: 131072

2. count_tokens

Counts the number of tokens in a string for a specific model. It loads the correct tokenizer to give you an accurate count.

from localgrid import count_tokens

text = "This is a test sentence for my model."
model = "google/gemma-3-12b"#lm-studio format

token_count = count_tokens(text, model)

print(f"The text has {token_count} tokens according to {model}.")
# The text has 9 tokens according to google/gemma-3-12b.

3. preload_tokenizers (optional async)

Due to the decision to keep this package fully offline capable there is a little delay when initially loading tokenizers from disk.

If you're using this in a server and want to avoid a tiny delay on the first call, you can preload the tokenizers into memory when your app starts.

import asyncio
from localgrid import preload_tokenizers

async def main():
    # Preloads all 25 base tokenizers
    await preload_tokenizers()
    
    # Or just preload specific ones
    await preload_tokenizers(families=["llama", "gemma"])

if __name__ == "__main__":
    asyncio.run(main())

Licensing & Included Tokenizers

All included tokenizer files are bundled with their original licenses (ex: LICENSE and tokenizer_config.json). You can find these within the installed package.

Note: Due to licensing restrictions, the following tokenizer families are not included in this package:

• command-r
• stablelm2
• codestral

As a result they default to generic tokenizers

If someone knows more about me than this please let me know and I will add them ASAP