# Conflicts: # studio/backend/core/inference/diffusion.py # studio/frontend/src/features/images/images-page.tsx
471 B
471 B
- Decision: attributed Z-Image diffusers resident-GPU
defaultgains almost entirely tocompile_repeated_blocks;channels_last/cudnn.benchmarkkept because free no-ops on DiT but useful for UNet. - Evidence:
compile_probe.pycompile-only eager1.83 → 0.79 s+56.6%;perf_verify.pyfull default1.80 → 0.82 s-54.7%. - Not isolated:
inference_mode, TF32, fused QKV; offered clean ablation onGPU 6. - Completed: answered attribution and caveats.