33x Speedup with RTX 4070 | Generated by AI

Home 2026.07 English 中文

---

Done! Here’s the full transformation:

---

🚀 Before vs After (RTX 4070)

Stage	CPU (before)	GPU (now)	Speedup
Text encoding	2.2s	0.3s	7×
Flux diffusion (4 steps)	~487s	12.8s	38×
VAE decode	37s	2.7s	13×
Total	526s	~16s	33×

What was done

1. Rebuilt sd_cpp with CUDA — cmake -DSD_CUDA=ON detected your RTX 4070 (compute 8.9)
2. Stopped llama-server to free ~6 GB VRAM
3. Key optimizations in run.sh:
   • --backend diffusion=cuda,clip=cpu,vae=cuda,t5xxl=cpu — Flux + VAE on GPU, text encoders on CPU
   • --vae-tiling — VAE compute buffer dropped from 6.6 GB → 416 MB with minimal speed loss
   • --max-vram 10 — leaves headroom for desktop apps

Issues found

• T5XXL file t5xxl_fp16.safetensors has corrupt tensors (“data offsets out of bounds”). Without it, prompt adherence is degraded. Re-download from here to fix.

Usage

cd /mnt/data/zz/flux
./run.sh "your prompt here"

Output saved to outputs/flux_<timestamp>.png (~1.7 MB).

---

Generated by AI. Curating and sharing still takes effort. If you find it useful, feel free to donate. WeChat: @lzwjava · X: @lzwjava · Say hi 👋

Back Donate