leejet
[Bug] Z-Image render time
Git commit
Operating System & Version
Ubuntu 24.04.3
GGML backends
Vulkan
Command-line arguments used
./sd --diffusion-model ./models/z_image_turbo-Q5_0.gguf --vae ./models/ae.sft --llm ./models/Qwen3-4B-Instruct-2507-Q4_K_M.gguf -v -p "a mix of anime and 3D style, a beautiful girl, perfect face, detailed face, walking in cyberpunk city at night, futuristic dress that emite light, intricate details, luxurious background, neon glow, highly detailed, 4K, professional photographie" --seed -1 --steps 9 --cfg-scale 1.0 --offload-to-cpu --diffusion-fa -H 512 -W 768
Steps to reproduce
It is random.
What you expected to happen
8 to 9s/it
What actually happened
Sometime, it goes up to 50s/it. And if launch the exact same command, it go back to 8 to 9s/it
Logs / error messages / stack trace
No response
Additional context / environment details
No response
Sometime, it goes up to 50s/it. And if launch the exact same command, it go back to 8 to 9s/it
Are you low on VRAM? Memory pressure could cause this (see ggml-org/llama.cpp#17605 for details, and a possible workaround).
I have 8Gb or VRAM.
Edit: I just checked with nvtop, sd.cpp only uses 5GB of VRAM out of the 8GB available.