AINanoGPT Speedrun world record reportedly drops to 68.0 seconds
A post reports a 5.8-second improvement, with changes including a smaller query/key dimension, fused Triton operations and more FP8 computation.
TLDR
A post reports a new NanoGPT Speedrun world record of 68.0 seconds, a 5.8-second improvement. It describes changes that reduce the query/key dimension from 128 to 96, combine several operations in a Triton kernel and move selected computations to FP8, an 8-bit floating-point format. The post describes a pull request involving more than 1,000 lines of Triton and four different FP8 scaling protocols.
Combined views
3.6K
2 Sources, first seen 12h ago
NanoGPT Speedrun world record reportedly drops to 68.0 seconds
A post reports a 5.8-second improvement, with changes including a smaller query/key dimension, fused Triton operations and more FP8 computation.
