[Bug] ggml commit cc937ae causes Krea2 generation to OOM on AMD dGPU #2254

Git commit

stable-diffusion.cpp at c6dfa9d and ggml at c57b71a

Operating System & Version

Linux 7.3.0-gc339f4010dea-dirty

GGML backends

Vulkan

Command-line arguments used

sd-cli -M img_gen --diffusion-model /home/max/ComfyUI/models/diffusion_models/kroma-0.3.1-turbo-opd-Q8_0.gguf --llm /home/max/ComfyUI/models/text_encoders/qwen-3-vl-4B-Q8_0.gguf --vae /home/max/ComfyUI/models/vae/qwen-BF16.safetensors --lora-model-dir /home/max/ComfyUI/models/loras/ -p "$(cat /home/media/Samsung_SSD_970_Evo/text/machines/prompt_pos.txt)" --conv-direct on --flash-attn on --mmap --offload-to-cpu --rng cpu --cfg-scale 1 --sampling-method euler --scheduler discrete --steps 8 -H 1184 -W 888 -s 614094079319347 -o (date --rfc-3339=second).png

Steps to reproduce

I compiled stable.cpp at c6dfa9d and its ggml fork at c57b71a, and have observed warnings, errors and failures during image generation right away.
These were somewhat non-deterministic between different image generations of the same workflow and the same commit, but are definitely attributable to and introduced by cc937ae.

I managed to make this a lot more reproducible by running Minecraft in the background, which used about 2GiB of VRAM.
With this, running my workflow at the specifiec commits have always resulted in a failure quite early, at the memory mapping stage or the first few sampling steps:

  |==================>                               | 3/8 -  11.16s/it - ETA 01:06[W] model manager memory on Vulkan0: reported free 117.54 MB / total 12272.00 MB, tracked weights 1462.25 MB / other runtime 0.00 MB / current runtime 1296.77 MB --- model_manager.cpp:1794
[W] model manager cannot make enough memory available on Vulkan0: need 512.00 MB device / 1296.77 MB budget, available 117.54 MB device / unlimited budget --- model_manager.cpp:2024
[E] krea2 segment 22/29 (krea2.blocks.21) failed during allocated capacity check --- ggml_runner.cpp:933
[E] diffusion model compute failed --- diffusion_engine.cpp:2655
[E] Diffusion model sampling failed --- diffusion_engine.cpp:2888
[E] sampling for image 1/1 failed after 48.15s --- image.cpp:961
[E] generate failed --- main.cpp:926

However, running stable.cpp with ggml at 6a89493 and with Minecraft in the background, did not result in a single warning, error or failure on multiple runs of the same workflow.

Prepending GGML_VK_REUSE_BUFFERS=0, introduced by cc937ae, did not help.

What you expected to happen

The opposite of the above.

What actually happened

See above.

Logs / error messages / stack trace

No response

Additional context / environment details

apiVersion        = 1.4.363 (4211051)
driverVersion     = 26.99.99 (109457507)
vendorID          = 0x1002
deviceID          = 0x73df
deviceType        = PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
deviceName        = AMD Radeon RX 6700 XT (RADV NAVI22)