[Shiunbird]

Mrs Rossignol got a LLM to run on a PowerBook G4. The single thread delta between his 12 inch G4 and a modern computer is much smaller than I thought. As a bonus, he optimises part of the code for Altivec.

http://www.theresistornetwork.com/2025/0...-llms.html

Any of our mad geniuses to get it to run under IRIX?
I gave it a shot during the weekend, but I couldn't get it to build even under modern Linux.

[jirka]

Any of our mad geniuses to get it to run under IRIX?
I gave it a shot during the weekend, but I couldn't get it to build even under modern Linux.

Well, it works under the Fedora on the POWER9 (ppc64le). On the IRIX it compiles. I had to comment out the function time_in_ms() (it is probably fixable but I was lazy) and then it compiled wihtui errors (gcc 4.7 from Nekoware fith a flag  -std=c99).

But it ends with "malloc failed!" even for the smallest available model. My machine is SGI O2 with 768MB of RAM and 128MB of swap.

[Shiunbird]

How did you get it to compile?
I am the most stupid in this kind of thing and, unless I write the thing myself from scratch, I hardly ever manage to port anything. I got stuck at trying to figure out the makefile. =(

I wonder if it would run on my Octane with 1.5GB. I am very curious about the performance.

Recently, I was running a few tests with Shake, tried my Octane vs a G5 vs my Intel Mac Pro and the single thread performance difference is much smaller than I thought. Of course, once you let it run with 8 threads (Shake's maximum), it obliterates the rest.