Dear all,
Recently I purchased "[Build a Large Language Model (From Scratch)](https://www.manning.com/books/build-a-large-language-model-from-scratch)" by Sebastian Raschka, so that I could learn more about how to build and/or fine-tune a LLM, and even developing some applications with them. I have also been skimming and reading on this sub for several months, and have witnessed many interesting developments that I would like to follow and experiment with.
However, there is a problem: The machine I have is a very old Macbook Pro from 2011 and I probably would not be able to afford a new one until I'm in graduate school next year. So I was wondering that, other than getting a new machine, what are the other (online/cloud) alternatives and/or options that I could use, to experiments with LLMs?
Many thanks!
I've never used it, but I think Google Colab has a free plan.
As another option, you can rent a machine with a decent GPU on vast.ai. An Nvidia 3090 can be rented for about $0.20/hr.
I am on a 2019 mac and finding it difficult too.
Best bet would be to start with a small language model?
Make yourself comfortable with
https://blogs.oracle.com/database/post/freedom-to-build-anno...
https://gist.github.com/rssnyder/51e3cfedd730e7dd5f4a816143b...
https://www.reddit.com/r/oraclecloud/
or any other offer.
Deploy some minimal Linux on them, or use what's offered.
Plus optionally, if you don't want to instantly start coding from first principles/scratch, make use of established and excellent solutions, like
https://future.mozilla.org/builders/news_insights/introducin...
https://ai-guide.future.mozilla.org/content/running-llms-loc...
https://github.com/mozilla-Ocho/llamafile
https://justine.lol/matmul/
and parallelize them with
https://github.com/b4rtaz/distributed-llama
Obviously this needs some knowledge of the command line, so get a good terminal emulator like
https://iterm2.com/
Mend, bend, rend that stuff and see what works how and why, and what not.
Edit: Optionally, if you really want to go low-level, with some debugger like
https://justine.lol/blinkenlights/
for 'toy-installations' of smallest models.
'Toy' because that doesn't fully support the CPU-instructions which are used in production.
Could still help conceptually.