{"slug": "the-lr-death-spiral-muon-dion3-debugging-saga", "title": "The LR Death Spiral 🧨 — Muon/Dion3 Debugging Saga", "summary": "An engineer preparing optimizer experiments for an upcoming PyTorch Conference talk found that the Dion3 implementation in OLMo-core broke training with a learning rate death spiral while testing AdamW, Muon, and Dion. The debugging trail traced the failure to distributed optimizer state, LR scheduling, and a PyTorch tensor aliasing trap, which the engineer documented in a write-up. More optimizer experiments are planned before PyTorch Con.", "body_md": "I was supposed to be running optimizer experiments for my upcoming PyTorch Conference talk.\n\nInstead, my learning rate entered a death spiral.\n\nI was testing AdamW, Muon, Dion, and the latest Dion3 implementation in OLMo-core—and Dion3 broke things in some very interesting ways.\n\nTurns out the debugging trail involved distributed optimizer state, LR scheduling, and a particularly nasty PyTorch tensor aliasing trap.\n\nA fun debugging experience nonetheless.\n\nI wrote up the whole adventure here:\n\nMore optimizer experiments coming before PyTorch Con!", "url": "https://wpnews.pro/news/the-lr-death-spiral-muon-dion3-debugging-saga", "canonical_source": "https://discuss.huggingface.co/t/the-lr-death-spiral-muon-dion3-debugging-saga/180675#post_1", "published_at": "2026-09-22 01:42:49+00:00", "updated_at": "2026-09-22 01:52:37.194733+00:00", "lang": "en", "topics": ["machine-learning", "ai-research", "developer-tools"], "entities": ["PyTorch Conference", "AdamW", "Muon", "Dion", "Dion3", "OLMo-core", "PyTorch"], "alternates": {"html": "https://wpnews.pro/news/the-lr-death-spiral-muon-dion3-debugging-saga", "markdown": "https://wpnews.pro/news/the-lr-death-spiral-muon-dion3-debugging-saga.md", "text": "https://wpnews.pro/news/the-lr-death-spiral-muon-dion3-debugging-saga.txt", "jsonld": "https://wpnews.pro/news/the-lr-death-spiral-muon-dion3-debugging-saga.jsonld"}}