Stochastic rounding lets BF16 RL post-training converge as fast as FP16
heghbalz · x · 2026-11-13
A technical thread on RL post-training numerics: casting FP32 shadow weights to BF16 with stochastic rounding lets BF16 RL post-training converge as fast as FP16. At the same learning rate, vanilla BF16 lags — even when training and inference backends produce bit-exact logprobs.
Key points:
- The bottleneck for low-precision RL training is rounding error in weight updates, not backend mismatch
- Stochastic rounding makes BF16 viable for RL post-training, cutting memory/bandwidth costs
- Directly useful for engineering teams running large-scale RL training
More from Infra
- Anthropic's Q3: revenue up 47% but compute costs jumped 77%, swinging to an operating loss — bigdata · 2026-11-13
- Details of Anthropic's self-hosted Managed Agents beta and its security-perimeter architecture — ClaudeDevs · 2026-11-13
- Anthropic launches self-hosted Claude Managed Agents in private beta for on-prem deployment — ClaudeDevs · 2026-11-13
- SpaceX commits $220M for water infrastructure to support its $24B AI supercomputing site — Polymarket · 2026-11-13
- SpaceX commits $220M to Texas water infrastructure for its $24B AI supercomputing site — Polymarket · 2026-11-13
- vLLM Semantic Router Istio policy bypass: one header skips the security filter — colinmcnamara · 2026-11-13