Show HN: How to Speed Up VLA Training Step by Step:3× Acceleration on Pi0.5
A step-by-step account of accelerating Pi0.5 SFT on LoongForge: empty camera-slot pruning, partial torch.compile, FlexAttention, selective recomputation, layered compilation to hide gradient AllReduce, and TorchAO FP8. 2.63x the community throughput in bf16 and 3.00x in FP8.