Drain the straggler
Fabric 07Fix the board. Turn every goal green.
Goals
Not tested yet
Eight H100 train an 8B model with data parallel; a ninth GPU is a hot spare on the same switch. Each step ends with a ring all-reduce, so every GPU waits for the slowest. A step took 1.55 s last week. Bring the median step back under 1.75 s with one change.
03:10 Step time 2.15 s, was 1.55 s. Graph: kernel time per GPU; gpu 5 is 1.6 times the others.
2 GPUs · batch 64 · job 1.5 h · 2.1 USD