vime × RL-Kernel × AMD: Bitwise Train–Rollout Consistency on ROCm
vime × RL-Kernel × AMD: Bitwise Train–Rollout Consistency on ROCm
摘要
vime and RL-Kernel align selected-token logprobs bit for bit across Megatron training and vLLM rollout on AMD Instinct MI300X, with zero mismatches across 200 GRPO steps.
vime and RL-Kernel align selected-token logprobs bit for bit across Megatron training and vLLM rollout on AMD Instinct MI300X, with zero mismatches across 200 GRPO steps.
译自 vLLM · 官方博客 · 录于 二〇二六年九月二十四日