[Bug] multi-device splitting failing on Unet models

[wbruna]

Git commit

b11c95a (master-749-b11c95a)

Operating System & Version

Debian 13

GGML backends

Vulkan

Command-line arguments used

sd-cli -m dreamshaper_8.safetensors --vae-conv-direct -o flower01.png -p flower --backend 'Vulkan0&Vulkan1'

Steps to reproduce

Just tried multi-device splitting:

Vulkan0	Vulkan1	Vulkan0&Vulkan1

The output suggests that Unet models do not benefit that much:

[WARN ] layer_split_partition.cpp:108 - Conditioner model: no transformer blocks found for a layer split; keeping tensors on compatible backends starting from Vulkan0
[INFO ] layer_split_partition.cpp:183 - Diffusion model layer split: Vulkan0 <- blocks [0, 2) + non-block tensors
[INFO ] layer_split_partition.cpp:183 - Diffusion model layer split: Vulkan1 <- blocks [2, 12)
[INFO ] stable-diffusion.cpp:1247 - using VAE for encoding / decoding
[INFO ] auto_encoder_kl.hpp:527 - vae decoder: ch = 128
[WARN ] stable-diffusion.cpp:316 - vae module does not support multiple runtime backends; using Vulkan0

However, there is no indication that it isn't supported.

The backend itself works: Anima produces correct results.

What you expected to happen

It should either produce coherent outputs, ignore the splitting, or error out.

What actually happened

Bad generated images.

Logs / error messages / stack trace

No response

Additional context / environment details

No response