Llama Chat and Code Llama are good at coding – InfoWorld
By Martin Heller
Contributor, InfoWorld |
A funny thing happened when I was playing with Poe, a chatbot aggregator from Quora. I selected the recently released free almost-open-source Llama 2 70B Chat model from Meta and gave it the prompt “Generate a Python program to scrape a website. Include tests for python.org and infoworld.com.” That’s a shorter version of a prompt I used to test CodeWhisperer, Bard, and Copilot X in June. None of those three models generated a working program.
To my utter surprise, Llama 2 70B Chat aced this test:
The bot went on to suggest some follow-up questions:
The generated code looked good to me, so I copied it into Visual Studio Code, saved it, and ran it. It ran perfectly:
Comparing the Llama-generated code with the CodeWhisperer-generated code, the major difference is that Llama used the html.parser model for Beautiful Soup, which worked, while CodeWhisperer used the lxml model, which choked.
I also asked Llama 2 70B Chat to explain the same sample program I had given to CodeWhisperer, Bard, and Copilot X. CodeWhisperer doesn’t currently have a chat window, so it doesn’t do code explanations, but Bard did a great job on this task and Copilot X did a good job.
Llama’s explanation (shown above) is as good, or possibly better, than what Bard generated. I don’t completely understand why Llama stopped in item 12, but I suspect that it may have hit a token limit, unless I accidentally hit the “stop” button in Poe and didn’t notice.
For more about Llama 2 in general, including discussion of its potential copyright violations and whether it’s open source or not, see “What is Llama 2? Meta’s large language model explained.”
A couple of days after I finished working with Llama 2, Meta AI released several Code Llama models. A few days after that, at Google Cloud Next 2023, Google announced that they were hosting Code Llama models (among many others) in the new Vertex AI Model Garden. Additionally, Perplexity made one of the Code Llama models available online, along with three sizes of Llama 2 Chat.
So there were several ways to run Code Llama at the time I was writing this article. It’s likely that there will be several more, and several code editor integrations, in the next months.
Poe didn’t host any Code Llama models when I first tried it, but during the course of writing this article Quora added Code Llama 7B, 13B, and 34B to Poe’s repertoire. Unfortunately, all three models gave me the dreaded “Unable to reach Poe” error message, which I interpret to mean that the model’s endpoint is busy or not yet connected. The following day, Poe updated, and running the Code Llama 34B model worked:
As you can see from the screenshot, Code Llama 34B went one better than Llama 2 and generated programs using both Beautiful Soup and Scrapy.
Perplexity is website that hosts a Code Llama model, as well as several other generative AI models from various companies. I tried the Code Llama 34B Instruct model, optimized for multi-turn code generation, on the Python code-generation task for website scraping:
As far as it went, this wasn’t a bad response. I know that the requests.get() method and bs4 with the html.parser engine work for the two sites I suggested for tests, and finding all the links and printing their HREF tags is a good start on processing. A very quick code inspection suggested something obvious was missing, however:
Now this looks more like a command-line utility, but different functionality is now missing. I would have preferred a functional form, but I said “program” rather than “function” when I made the request, so I’ll give the model a pass. On the other hand, the program as it stands will report undefined functions when compiled.
Returning JSON wasn’t really what I had in mind, but for the purposes of testing the model I’ve probably gone far enough.
At Google Cloud Next 2023, Google Cloud announced that new additions to Google Cloud Vertex AI’s Model Garden include Llama 2 and Code Llama from Meta, and published a Colab Enterprise notebook that lets you deploy pre-trained Code Llama models with vLLM with the best available serving throughput.
If you need to use a Llama 2 or Code Llama model for less than a day, you can do so for free, and even run it on a GPU. Use Colab. If you know how, it’s easy. If you don’t, search for “run code llama on colab” and you’ll see a full page of explanations, including lots of YouTube videos and blog posts on the subject. Note that while Colab is free but time-limited and resource-limited, Colab Enterprise costs money but isn’t limited.
If you want to create a website for running LLMs, you can use the same vLLM library as used in the Google Cloud Colab Notebook to set up an API. Ideally, you’ll set it up on a server with a GPU big enough to hold the model you want to use, but that isn’t totally necessary: You can get by with something like a M1 or M2 Macintosh as long as it has enough RAM to run your model. You can also use LangChain for this, at the cost of writing or copying a few lines of code.
If you are using an Arm-based Macintosh as your workstation, you can run Llama models locally as a command-line utility. The invaluable Sharon Machlis explains how to use Ollama; it’s easy, although if you don’t have enough RAM memory for the model it’ll use virtual memory (i.e. SSD or, heaven forfend, spinning disk) and run really slow. (Linux and Windows support is planned for Ollama.)
I tried out Ollama with several models (of the many it supports) on my M1 MacBook Pro, which unfortunately has only 8GB of RAM. I started with my standard Python web-scraping code generation task using Llama 2, apparently one of the smaller models (7B?). The result is similar to what I got from the Llama 2 70B model running on Poe, although not as well-structured. Note that Ollama only downloads the model the first time it needs it.
With that baseline established, I tried the same prompt using Code Llama. Again, I didn’t specify the model size, but it looks like it is 7B.
Llama 2 Chat can generate and explain Python code quite well, right out of the box. There’s no need to fine-tune it further on code generation tasks.
Code Llama’s nine fine-tuned models offer additional capabilities for code generation, and the Python-specific versions seem to know something about Python classes and testing modules as well as about functional Python.
Copyright © 2023 IDG Communications, Inc.
