Diffusers Tensor Parallel Loading Gets Major Speedup

Hugging Face Diffusers upgraded tensor parallel loading, cutting Flux.2-Dev load times by 2.4x with TP=4 on A10G GPUs. The core PR enables shard-wise loading during .frompretrained(), saving about 89% CPU memory.

2026-10-01 ~ 2026-10-01 · 2 related posts

• Diffusers tensor parallel loading gets 2.4x faster, cuts CPU memory by 89% — RisingSayak · 2026-10-01
• Diffusers PR #14544: shard tensor-parallel checkpoints on load and save — RisingSayak · 2026-10-01