[Bug] regression: increased resident memory usage with mmap since master-691-563137a

[wbruna]

Git commit

563137a (master-691-563137a)

Operating System & Version

Debian 13

GGML backends

Vulkan, HIP

Command-line arguments used

sd-cli --backend ROCm0 -p 'sunset' --diffusion-model z_image_turbo-Q8_0.gguf --llm Qwen3-4B-UD-Q4_K_XL.gguf --vae ae_bf16.safetensors --cfg-scale 1 --steps 9 -W 1024 -H 1024 --fa --mmap --offload-to-cpu

Steps to reproduce

Since master-691-563137a, offloading with mmap is using a lot of resident memory, instead of the expected shared memory. As measured by top during inference:

	RES	SHR
master-690	7,7g	<1.0G
master-690 --mmap	10.0G	9.4G
master-691	7,6g	<1.0G
master-691 --mmap	16.1G	<1.0G

(note that it's expected that RES shows up larger with mmap, since it counts the memory-mapped areas too, and those are never released from memory; but it shouldn't result in a plain increase of resident memory usage)

The log keeps showing mmap being used for the models:

[INFO ] stable-diffusion.cpp:491  - Version: Z-Image 
[DEBUG] model_loader.cpp:813  - using mmap for I/O
[INFO ] model_loader.cpp:819  - using mmap for 'z_image_turbo-Q8_0.gguf'
[DEBUG] model_loader.cpp:813  - using mmap for I/O
[INFO ] model_loader.cpp:819  - using mmap for 'Qwen3-4B-UD-Q4_K_XL.gguf'
[DEBUG] model_loader.cpp:813  - using mmap for I/O
[INFO ] model_loader.cpp:819  - using mmap for 'ae_bf16.safetensors'

Later releases show different values, but similar behavior: increased RES usage with low shared memory.

What you expected to happen

shared memory being used, to reduce RAM pressure

What actually happened

increased RAM usage

Logs / error messages / stack trace

No response

Additional context / environment details

No response

[dhruvil-codes]

Hey, would love to work on this bug!

[wbruna]

This depends on the backend: I see the resident memory spike on ROCm, but Vulkan keeps using shared memory as expected.

[sz1kormar]

I also noticed the same problem on Vulkan with Windows 10.