Notes on Pretraining Parallelisms and Failed Training Runs

Notes on pretraining parallelisms and failed training runs.

Dwarkesh documents pretraining failure modes and parallelism tradeoffs: expert choice and token dropping can break causality in MoE routing, FP16 collectives can bias repeated additions after values exceed 1024, pretraining FLOPs are given as 6ND, B300 HBM is listed as 288GB, and FSDP communication can reach params × 3 with reduce-scatter.

Why it matters: HKR-H/K/R all pass: Dwarkesh’s notes expose concrete pretraining failure modes and numbers. The systems-training focus is specialized, so it sits in the high-quality band rather than same-day must-write.