I immediately opened the package and installed the WD 2TB 2230 model.
I tested the code I had been writing over the weekend, and it's great because of the large memory (although...it's still not enough...).
At first, I ran it with the Qwen 3.6 27B model (because it was considered the smartest...), but...
the fine-tuning time was incredibly long, so I wandered around trying different models.
Ultimately, I realized that a 200k context was insufficient for my intended use, so I settled on the Qwen/Qwen2.5-7B-Instruct-1M model.
For processing 1 million-token sequences:
After roughly 100 case studies, it seems to be working well for now.
Of course, 90% of the time required for each execution is spent on program execution and weight loading,
and the results come out almost instantly. Cool...
It seems like I'm about to embark on a journey to find the sweet spot where I can achieve the desired results by increasing the learning numbers.
I need to update the interface code, which is still under development...
My current dilemma is whether to officially announce this project and make it a funded endeavor (assuming it's feasible)
or just keep working on it secretly while earning money. Haha.
Since it's for work, I brought it to the office (because my electricity bill at home is precious...).
I placed it on top of a heatsink.
