Legacy FSDP logic hinders operator fusion and slows training

SeunghyunSEO7 · x · 2026-08-24

Discusses how legacy logic from FSDP1, still present in FSDP2, intentionally skips installing guards. This creates more symbolic arguments, making it harder to fuse elementwise operations like LayerNorm and RoPE, thereby slowing down training.

Related event: torch.compile with FSDP Dynamic Shapes Degrades MFU(3 posts)→

Original post →

More from Infra

Infra channel →