On a single LLaMA2-7B MMLU run, updating the upper 24 of 32 layers with PaCA gave 52.02 average accuracy versus LoRA's 52.15, while cutting training time from 7:17 to 5:15.
Accelerate gpu concurrent kernel execution by mitigating memory pipeline stalls
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
support 1representative citing papers
citing papers explorer
-
LoRA Is Slower Than You Think
On a single LLaMA2-7B MMLU run, updating the upper 24 of 32 layers with PaCA gave 52.02 average accuracy versus LoRA's 52.15, while cutting training time from 7:17 to 5:15.