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.

Original post →

More from Infra

Infra channel →