Dear authors,

Thanks for the great help. Currently we've tried to measure SDXL base model.
I've tried (win-vulkan)

sd-cli.exe -m sd_xl_diffusion_base_x1.safetensors -W 1024 -H 1024 -p "a lovely cat" --steps 7 -v

I have two computers:

C1: Lunar Lake, CPU: Intel Ultra 268V, GPU: Intel R AI Boost, RAM: 32G
C2: Panther Lake, RAM: 64G

However, for C1, the image cannot be generated. The final log message shows
[DEBUG] ggml_extend.hpp - vae compute buffer size: 2560.25 MB (VRAM)

If I enable "--vae-tiling --vae-conv-direct", the VRAM size will drop to 176MB and it can execute successfully.

How could we know if the platform can work without setting --vae-tiling ?

I've tried to dump VulkanInfo and observe "maxStorageBufferRange" and "memoryHeaps[0]". But actually I have another platform which can run without setting --vae-tiling. The values ("maxStorageBufferRange" and "memoryHeaps[0]") of the two computers are almost the same.

Thanks