ACRL:訓練-推論エンジン乖離の適応制御でFP8量子化下のRL学習を安定化Huawei's ACRL framework monitors the discrepancy between training…
AI要約HuaweiのACRLは、LLMのRL学習でFSDP/vLLM間の精度差(BF16 vs FP8)による実質的なoff-policy化を適応的に補正し、トークン単位の勾配重み調整でBF16基線を上回る精度をわずか0.1%のオーバーヘッドで実現する。
AI SUMMARYHuawei's ACRL framework monitors the discrepancy between training (FSDP/Megatron) and inference (vLLM/SGLang) engines caused by FP8 vs BF16 precision gaps, then adjusts per-token gradient weights to prevent training collapse while outperforming BF16 baselines across 3B–32B Dense and MoE models.
og
