vLLM Unveils IsoExec to Eliminate Trainer-Inference Mismatch in SkyRL
#vllm#sky-rl#reinforcement-learning#execution-layer
vLLM announced IsoExec, a unified execution layer for SkyRL pipelines that aligns numerical behavior between the vLLM and Megatron runtimes. On Qwen3.5-35B-A3B, it reduces the average rollout-versus-training logprob difference to below 1e-6, with a 25% overhead.
Coverage timeline
vLLM BlogAlexander Jiang and the SkyRL Team
IsoExec unifies numerical execution across SkyRL's vLLM and Megatron runtimes, reducing the average rollout-versus-training logprob difference below 1e-6 on Qwen3.5-35B-A3B with 25% overhead.