As the title suggests, basically i have a few LLM models and wanted to see how they perform with different hardware (Cpus only instances, gpus - t4, v100, a100). Ideally it’s to get an idea on the performance and overall price(vm hourly rate/ efficiency)
Currently I’ve written a script to calculate ms per token, ram usage(memory profiler), total time taken.
Wanted to check if there are better methods or tools. Thanks!
I’ve seen a couple around, grabbed this one off my list:
lm-evaluation-harness
Thanks. Does this also conduct compute benchmarks too? Looks like this is more focused on model accuracy (if I’m not wrong)
seems like, keep an eye, when i run across one I will post it, usually to the model’s community.
sure, thank you!