Reading up on RTX 4090

1 deep · digging since mar 11

Show HN: How I topped the HuggingFace open LLM leaderboard on two gaming GPUs
Duplicating a block of ~7 middle transformer layers in Qwen2-72B, without weight changes, boosted benchmark scores to #1 on two RTX 4090s, suggesting pretraining carves discrete functional circuits.