SkyRL's IsoExec brings bitwise logprob parity between vLLM rollouts and Megatron training
vllm_project · x · 2026-08-21
In RL training, a vLLM rollout engine and a Megatron trainer can run the same policy yet disagree on token logprobs due to floating-point non-associativity.
SkyRL's IsoExec combines an execution contract with a unified model, aligning rounding-sensitive execution choices across rollout and training. Bitwise parity holds across different TP/EP/SP layouts; for Gated DeltaNet, a chunkwise-parallel recurrent algorithm makes parallel training/prefill bitwise identical to recurrent decode.
Results: Qwen3.5-35B-A3B + DAPO, 8×H100, 50 steps — logprob diff drops from 1.6e-2 to 6.7e-7, with 25.3% full-step overhead. vLLM's scheduler and CUDA graphs still apply.
More from Infra
- Google's Willow chip cuts quantum errors 40x, raising Q-Day concerns for Bitcoin — jamestagg · 2026-10-16
- Vibe Coding a Data Pipeline Is Easy; Running It at Billions of Rows Is Not — Zachly · 2026-10-16
- Large data centers employ only ~50-150 permanent on-site staff, Microsoft says — brezina · 2026-10-16
- Prediction: Opus 5.5-level models on laptops by 2027, backed by Qwen 27B data — cephaloform · 2026-10-16
- Oracle down 12% in a week, nearly 50% off June high as OpenAI exposure spooks market — alvelda · 2026-10-16
- US data center moratoriums top 400 — an AI bull explains why they're actually good for AI — kevinsxu · 2026-10-16