Thinky's Dimuon FAQ: no RL gains over AdamW, tune learning rate only
eliebakouch · x · 2026-10-15
Thinky's Dimuon optimizer FAQ answers key questions: in their experiments, Dimuon performs about the same as AdamW when doing RL without SFT — no improvement. Optimizer state resets between training stages unless you resume via createtrainingclientfromstatewithoptimizer. For Dimuon, they recommend tuning only the learning rate, since instability usually means it's too large; beta1 and gradclipnorm can be left alone. Author eliebakouch notes he was surprised it doesn't change RL.
More from Infra
- Tree packing lands in TRL AsyncGRPO, cutting agentic RL trainer compute 26x — SergioPaniego · 2026-10-15
- JSSA sparse attention hits 6.43x decoding throughput at 128K context on B200 GPUs — SNU-VLSI · 2026-10-15
- 189 benchmarks later: how to pick KV quants for Qwen3 27B on 16GB VRAM — Fancy-Snow7 · 2026-10-15
- Energy, compute, data are the only new currency: $15B Anthropic-SpaceX deal caps 22 days of AI shakeups — sierracatalina · 2026-10-15
- Titan RTX + 3090 without NVLink: is multi-GPU inference via Strata worth it? — Express_Quail_1493 · 2026-10-15
- ASML sold 17 EUV machines in Q3 at up to $400 million apiece — NinaDSchick · 2026-10-15