Notes on SOAP, Muon and KL-SOAP: AdamW hits a large-batch wall at 72B scale
tokenbender · x · 2026-07-24
- These notes summarize the paper “SOAP, Muon, and Beyond: Pushing LLM Pretraining Scales”.
- Main claim: the so-called critical batch size is not only a property of model and data; it is also an optimizer property, and AdamW breaks down when gradients get too clean at very large global batches.
- The author explains why synchronous pretraining at scale makes small batches communication-bound, and why large batches are economically attractive but expose optimizer limits.
- According to the notes, Muon and KL-SOAP beat AdamW at 72B scale, with KL-SOAP slightly ahead, while Muon uses less state than AdamW.
- The included diagrams argue that in MoEs, sparse experts may stay stable, but dense/shared parameters can become the bottleneck under huge batch sizes.
More from Research
- Diamandis: AI Cancer Navigator and $100 Genome Sequencing Close the Care Gap — PeterDiamandis · 2026-10-27
- EMNLP Oral: training-free Nexus Sampling cuts KV-cache memory 10x; OpenResearcher tops BrowseComp-Plus at 54.8% — yuz9yuz · 2026-10-27
- ICML 2026 position paper: LLM 'thinking traces' are a harmful anthropomorphism — rao2z · 2026-10-27
- Forking gap follows a power law in context frequency; masking rare n-grams shrinks it — ZimingLiu11 · 2026-10-27
- How forking happens: n-gram tables split logits into non-interacting subspaces — ZimingLiu11 · 2026-10-27
- The best LoRA explainer video yet: rank constraints meet intrinsic dimension — sophiamyang · 2026-10-27