We Tried ISO-AdamW. AdamW Kept Its Job.
DEV Community
We Tried ISO-AdamW. AdamW Kept Its Job.
A small GRPO experiment: ISO-AdamW scored 75.8% versus AdamW's 75.4% on GSM8K, with 47% more GPU memory. The idea, the 44x kernel detour, and what four answers mean.
0 comments
No comments yet.