Question about fine tuning a small model: I was burning cash on cloud GPUs until a Discord user set me straight
I used to spin up a big cloud instance every time I wanted to tweak a 7B model, and my bill hit $412 in one month back in March. Then a guy in a Discord server asked why I was not just running the 4-bit version on my own 3090 at home. I tried it out of spite and the whole fine tune finished in about 90 minutes, same loss curve as the cloud runs. Turns out I was renting power I already had sitting under my desk this whole time, which honestly stung a bit. Anyone else switch their whole workflow to local hardware after seeing the numbers, or do you still pay for cloud when a job gets big?
Watched the same thing happen with my buddy who was paying $60 a month for a storage unit while his garage sat half empty. People just get used to paying for things and stop asking if they still need to. I moved my whole setup local back in January and the only time I touch cloud now is when I need way more VRAM than my card has. Funny how we trust a random Discord guy more than our own hardware sometimes.