Baidu Intelligent Cloud Tech Hub

Aug 21, 2026 · Artificial Intelligence

Achieving 2.37× Faster OpenPI PyTorch Training with RLinf and Baidu Baige

By first aligning Pi0.5’s precision between JAX and PyTorch and then applying a full‑stack AI Infra overhaul—including data‑fetch slot pruning, uint8 H2D transfers, torch.compile fusion, per‑block FSDP restructuring and communication prefetch—the team raised OpenPI PyTorch training throughput from 74.2 sps to 175.6 sps (2.37×) while preserving accuracy and achieving over 91% scaling efficiency on 256 GPUs.