Development

My Best CoPilot Alternative - Running LLM on Local Machine

I’m always on the lookout for new and innovative tools to help me improve my coding skills and increase my productivity. Recently, I stumbled upon Code Llama, a free, open-source large lauguage model (LLM) developed by Meta that allows you to set up on your low-cost gaming desktop. In this blog post, I’ll be sharing my experience with Code Llama and how it can serve as a great alternative to GitHub CoPilot.

Setup Code Llama with ollama

ollama (https://ollama.ai/), which uses a Dockerfile-like configuration file. It also manage Docker layers from LLM to system prompt. ollama helps a lot to setup and run Code Llama. Although the website says Windows is coming soon, I follow steps in https://github.com/jmorganca/ollama and running it successfully. The following are my setup,

  1. Windows 11
  2. WSL 2 with Ubuntu
  3. Nvidia 3060 12GB
  4. https://github.com/jmorganca/ollama

If you do not have enough video RAM, it falls back to system RAM and CPU.

There is many model for you to choose (https://ollama.ai/library) but you should first try Code Llama. You can pull the Code Llama with ollama pull codellama like docker image. However, there are license agreemenet to accept (https://ai.meta.com/resources/models-and-libraries/llama-downloads/). Once you have request, accept and approve, you can start using it. If you do not there are many others library you can try.

❓ Why I choose 12GB Video RAM on display card instead of 8GB, 16GB, 24GB?
8GB video cards are more common and cheaper but with additional 4GB of video RAM (VRAM) you can run a large model in next tier. LLMs usually build in 3 tier with different model parameter size, each tier use certain amount of VRAM approxamiely. Below shows 12GB fits both 7B and 13B. RAM more than 12GB is a waste. Unless you spend time on quantizing model for next tier and accept the model runs much slower after quantization. Code Llama provides a 34B model that you can use with 24GB, which is the only use case for 24GB video card.

ModelSize in storageTypical memory usageVRAM 8GBVRAM 12GBVRAM 24GBVRAM 2 X 24GB
7B4GB7GB❌✅✅✅
13B8GB11GB❌❌✅✅
34B19GB23GB❌❌✅✅
70B40GB35GB❌❌❌✅

The largest video RAM size for consumer grade display card is 24GB so you will need to 2 video card.

How to Set Up Code Llama with Visual Studio Code

Setting up Code Llama with Visual Studio Code is easy and straightforward. Search and install an Visual Studio Code extension “Continue”. This extension allows you use LLM from the service provider and local LLM service like ollama. “Continue” starts an interactively tutorial and you should start to use it happily.

Local vs CoPilot (or anyother subscription base service)

There are many other CoPilot-like service provider that is free. Quality from free service usually poor so I will not compare on below table,

FeatureLocalCoPilot (or anyother subscription based service)
Cost of OwnershipUSD 200 for a second hand nVidia 3060 12GB. Excluding the PC deskop that I already have-
PriceFree. Electricity costs me around 1kWh a month, which is negligibleSubscription-based, starting from USD 10 a month
ROI20 months-
LLM for you to chooseMany. There are programming language specific model.Depends on the provider
Input token limit100k (Code Llama)1500 (Copilot)
SetupQuite easy with ollamaEasy
PrivacyYour data stays on your local machineData is sent to third party for processing

Input token is the “memory” of the context. GPT4’s input token is 8k thats why you feel the code generated is better than CoPilot when your requirements is complex. Please note the number could be inaccurate as technology is advancing quick.

It’s important to remember that the quality of the generated code is not just determined by the model used, but also by the system prompts and the context provided. This is because there are so many models you can use, and prompt could matter more than a model. You can customize system prompts that fits you best and switch between the model like hell from Continue. For example, below poor code could be generated from both Code Llama and GPT4 if you do not have a proper system prompt or even your context has been messed up,

No system prompt
write a python function to generate six unique numbers from 1 to 49

1
2
3
4
5
6
7
8
9
import random

def get_unique_numbers(n):
    nums = []
    while len(nums) < n:
        num = random.randint(1, 49)
        if num not in nums:
            nums.append(num)
    return nums

A simple system prompt makes a great different. Not to mention other parameters such as tempearture you can switch easily with ollama and Continue

System prompt: You are a seasoned programmer with a focus on using a single line of code to solve a problem.
Write a python function to generate six unique numbers from 1 to 49.

1
2
def generate_unique_numbers(n):
    return random.sample(range(1, n), k=6)

Conclusion

Security and privacy are my top concern. Running LLM is my best choice.

Share