[Performance] Optimize diffusion models: 18.6% average speedup via torch.compile, fused RoPE, and memory layout improvements#19623

Lyken17 wants to merge 1 commit intosgl-project:diffusion_skillssgl-project/sglang:diffusion_skillsfrom

Lyken17:diffusion_skillsLyken17/sglang:diffusion_skillsCopy head branch name to clipboard