Very good tutorial. You don't speach about the platform; can I assume it will work both Windows and Linux? Another thing: What's the recommended hardware configuration to install Llama 3 locally in our computers?
This was a great quick lesson. One thing I was seeing if anyone figured out, often I need to refer to very new documents on API etc, has anyone tied this into like. RAG structure, so we are always looking at latest document?
do we need both llama3:8b and instruct? can we not work only with instruct? Also I see your code works faster - could you specify your PC / system specs and config as it takes a good amount of time on my iMac 2017
The buttons don't do anything... note i'm working off line. The 4 buttons at the bottom of the add-ins panel just copy the code to the chat window. They don't do anything else and once clicked, the AI stops responding to questions. When i asked it what was wrong with the "explain selected code" the AI responded "nothing, its only meant to copy the code. Anyone know if this is broken for me or its simply an incomplete add-in...?
ollama : The term 'ollama' is not recognized as the name of a cmdlet, function, script file, or operable program. Check the spelling of the name, or if a path was included, verify that the path is correct and try again. At line:1 char:1 + ollama pull llama3:8b
SPOILER ALERT: this is not amazing, but you'll be able to make scrambled eggs on your laptop while it writes you a crud service that actually doesn't work
Does CodeGPT require me to be logged in? I'm all set up but if I ask it to explain something it just says "something went wrong! Try again.". Then i have to either quit and restart vscode or disable then enable the extension...
4 месяца назад
Thanks for sharing. I host the ollama server on a remote server. How do I make it connect to the remote machine instead of localhost?
This app looks like a good idea but its a long, long way from finished. Buttons (refactor, explain, document and fix bug in selected code don't do anything but copy selected code to the chat. If you use the clear button it clears the selected model etc. but not the history. I just asked it to write a basic api call for sveltekit and it wrote some pure garbage based on assuming the previous selection was part of the current question. I'm using a 2019 MBP with 32gb ram and its too slow to add any value so far... for me at least
I have 8 GB of VRAM and when autocomplete is on for Codyai copilot, my fans turn on full blast on my laptop. I have 64 GB of ram so it doesn't slow my pc down, but if it was running on your CPU and not your GPU it might slow your computer down. I don't think that it will slow your computer down if you have enough VRAM or a ton of ram, but it could depending on your computer's specs. There is also an extension called "Groqopilot" on vscode that requires that you supply it with a groq api key, and when you do it will create code for your lightning fast with llama3 70b which is of course the better model of llama3 8b. It doesn't autocomplete but it behaves very much like the tutorial we just watched.
Does anyone else get the feeling that the way AI's answer questions is based on the old Microsoft "clippy" assistant... annoyingly eager and can't answer anything much without wrapping it in a paragraph or so of irrelevance.... Very annoying to get 6 or 7 line answers where the only relevant bits are a number or a few words.
If you're using chat gpt you can change that in settings. I think in things like Ollama, you can also change your settings so that it get's straight to the point.
@@Fonzleberry I know thanks... just haven't had much luck though lol, at one point i got fed up and added an instruction to "only answer boolean questions with a yes or a no", I had to restart the model (bakllava) to get it to start answering properly again as it answered all questions with "yes" or "no". I don't get why the default mode is to burry all answers in information not requested. I guess someone redefined the word "conversational". Can't even ask whats 2+2 without an explanation lol
guys i installed it according to the vid but i cant run the ai and i saw somewhere that i need to put it in PATH but i dont know where the files are installed
You would need to know at least the basics of coding, and how an application is designed and structured. This writes the code for you but if you cannot read the code or at least understand what it's doing at a high level, then it's too early for you. It gives you 2/3 of the finished product. You just need to know how to integrate that code into your application. You need to know how to create an application, what are the different parts of an application, how to deploy and run an application.
I have the same issue. In the CodeGPT menu I only see as options "llama3:8b" and "llama3:70b", but not "llama3:latest" or "llama3:instruct", as I have them available (when I would go to a command line and do ollama list). When I select llama3:8b and enter a prompt, nothing happens. When I choose another model which I have installed, like "mistral" it works just fine...
Latency is pretty bad when im using llama3:70b on vscode for CodeGPT. I am on windows . I guess its with the underlying machine . Anything can be done here?
Get a better GPU. Preferably something with a lot of VRAM. If you’re attempting to do this on a laptop like many others here are, you’re setting yourself up for failure. If you’re serious about running large language models, don’t run them on a laptop.
I'm really not impressed with Llama:8B. I decided to skip Python, and go to Pascal. I asked to create a tic tac toe game, and have had nothing but problems with it. It CONSTANTLY forgets that Pascal is a declarative language and forgets to include the variable definitions, especially the loop variables. When i ask for it to revisit, this last time it decided to rewrite the function to draw the board in a console.log instead of a writeln. I mean, it rewrote the WHOLE function to be completely useless. I tried running the 70b, but the engine just kept prioritizing to my GTX970 instead of my RTX3070. The documentation on the site, as well as the Github repo just doesn't explain well enough where to put the weights on where the engine should calculate. I could pull the 970 out, but, meh.