[Bug] Every time i use vulkan it fails

[OnlineLearningTutorials]

Git commit

git rev-parse HEAD
be0e344

Every time when i use vulkan back-end it fails, while cpu mode works, while in llama.cpp i can use vulkan without an issue, after searching internet, i found there may vram issue, (my cpu and gpu uses unified ram in auto mode) but llama.cpp works perfectly in unified ram

ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon Graphics (RADV RENOIR) (radv) | uma: 1 | fp16: 1 | bf16: 0 | fp4: 0 | warp size: 64 | shared memory: 65536 | int dot: 0 | matrix cores: none
load_backend: loaded Vulkan backend from /opt/sd-master-0a565f2-bin-Linux-Ubuntu-24.04-x86_64-vulkan/libggml-vulkan.so
load_backend: loaded CPU backend from /opt/sd-master-0a565f2-bin-Linux-Ubuntu-24.04-x86_64-vulkan/libggml-cpu-haswell.so
[INFO ] stable-diffusion.cpp:720  - loading diffusion model from '/opt/sd-master-0a565f2-bin-Linux-Ubuntu-24.04-x86_64-vulkan/models/z_image_turbo/z_image_turbo-Q4_K_M.gguf'
[INFO ] model_loader.cpp:236  - load /opt/sd-master-0a565f2-bin-Linux-Ubuntu-24.04-x86_64-vulkan/models/z_image_turbo/z_image_turbo-Q4_K_M.gguf using gguf format
[INFO ] stable-diffusion.cpp:777  - loading llm from '/opt/sd-master-0a565f2-bin-Linux-Ubuntu-24.04-x86_64-vulkan/models/z_image_turbo/Qwen3-4B-Q4_K_M.gguf'
[INFO ] model_loader.cpp:236  - load /opt/sd-master-0a565f2-bin-Linux-Ubuntu-24.04-x86_64-vulkan/models/z_image_turbo/Qwen3-4B-Q4_K_M.gguf using gguf format
[INFO ] stable-diffusion.cpp:791  - loading vae from '/opt/sd-master-0a565f2-bin-Linux-Ubuntu-24.04-x86_64-vulkan/models/z_image_turbo/ae.safetensors'
[INFO ] model_loader.cpp:242  - load /opt/sd-master-0a565f2-bin-Linux-Ubuntu-24.04-x86_64-vulkan/models/z_image_turbo/ae.safetensors using safetensors format
[INFO ] stable-diffusion.cpp:908  - Version: Z-Image 
[INFO ] stable-diffusion.cpp:959  - Weight type stat:                      f32: 634  |     f16: 24   |    q4_K: 336  |    q5_K: 30   |    q6_K: 67   |    bf16: 4    
[INFO ] stable-diffusion.cpp:960  - Conditioner weight type stat:          f32: 145  |    q4_K: 216  |    q6_K: 37   
[INFO ] stable-diffusion.cpp:961  - Diffusion model weight type stat:      f32: 245  |     f16: 24   |    q4_K: 120  |    q5_K: 30   |    q6_K: 30   |    bf16: 4    
[INFO ] stable-diffusion.cpp:962  - VAE weight type stat:                  f32: 244  
[INFO ] stable-diffusion.cpp:1512 - using VAE for encoding / decoding
[INFO ] auto_encoder_kl.hpp:527  - vae decoder: ch = 128
[INFO ] stable-diffusion.cpp:1745 - total params memory size = 8467.58MB (VRAM 8467.58MB, RAM 0.00MB): text_encoders 3555.38MB(VRAM), diffusion_model 4752.20MB(VRAM), vae 160.00MB(VRAM), controlnet 0.00MB(N/A), extensions 0.00MB(N/A)
[INFO ] stable-diffusion.cpp:1864 - running in FLOW mode
[INFO ] stable-diffusion.cpp:5634 - generate_image 640x360
[INFO ] stable-diffusion.cpp:3174 - Using 'z_image_omni' preset for reference images
[INFO ] denoiser.hpp:1051 - get_sigmas with discrete scheduler
[INFO ] stable-diffusion.cpp:4386 - sampling using Euler method
  |##################################################| 386/386 - 787.94MB/s
[INFO ] model_loader.cpp:1304 - loading tensors completed, taking 2.94s (read: 1.19s, memcpy: 0.00s, convert: 0.07s, copy_to_backend: 0.30s)
[INFO ] stable-diffusion.cpp:5302 - get_learned_condition completed, taking 3.86s
[INFO ] stable-diffusion.cpp:5685 - generating image: 1/1 - seed 42
Terminated                 /opt/sd-master-0a565f2-bin-Linux-Ubuntu-24.04-x86_64-vulkan/sd-cli --diffusion-model /opt/sd-master-0a565f2-bin-Linux-Ubuntu-24.04-x86_64-vulkan/models/z_image_turbo/z_image_turbo-Q4_K_M.gguf --vae /opt/sd-master-0a565f2-bin-Linux-Ubuntu-24.04-x86_64-vulkan/models/z_image_turbo/ae.safetensors --llm /opt/sd-master-0a565f2-bin-Linux-Ubuntu-24.04-x86_64-vulkan/models/z_image_turbo/Qwen3-4B-Q4_K_M.gguf -p "create an image of solar system" -W 640 -H 360 --cfg-scale 1.0 --steps 5 -o output.png

Operating System & Version

Ubuntu 26.04

GGML backends

Vulkan

Command-line arguments used

/opt/sd-master-0a565f2-bin-Linux-Ubuntu-24.04-x86_64-vulkan/sd-cli --diffusion-model /opt/sd-master-0a565f2-bin-Linux-Ubuntu-24.04-x86_64-vulkan/models/z_image_turbo/z_image_turbo-Q4_K_M.gguf --vae /opt/sd-master-0a565f2-bin-Linux-Ubuntu-24.04-x86_64-vulkan/models/z_image_turbo/ae.safetensors --llm /opt/sd-master-0a565f2-bin-Linux-Ubuntu-24.04-x86_64-vulkan/models/z_image_turbo/Qwen3-4B-Q4_K_M.gguf -p "create an image of solar system" -W 640 -H 360 --cfg-scale 1.0 --steps 5 -o output.png

Steps to reproduce

use vulkan from release to produce image

What you expected to happen

vulkan back-end should work on gpu

What actually happened

gpu not used on cpu works

Logs / error messages / stack trace

No response

Additional context / environment details

No response

[farawayso]

从日志来看，核心线索是最后一行：

Terminated /opt/sd-master-0a565f2-bin-Linux-Ubuntu-24.04-x86_64-vulkan/sd-cli ...

"Terminated" 不是 sd.cpp 自己崩溃的，而是被操作系统（OOM Killer）强制杀掉的。

分析依据：

1. 设备是集成显卡 — AMD Radeon Graphics (RADV RENOIR)，这是 APU 核显，共享系统内存作为显存（UMA）。日志里 uma: 1 也证实了这一点。

2. 显存需求远大于可用资源 — 模型总参数占 8467 MB VRAM，但 RENOIR 核显的共享显存通常只有 2~4GB（系统会自动分配一部分），加上 fp16/bf16 混合精度，实际运行时还会叠加推理过程的中间张量，极易触发 OOM。

3. 正常崩溃会输出错误信息 — 比如 CUDA/Vulkan 错误、segmentation fault 等，但这里只有干净的 "Terminated"，这是 Linux 内核 OOM Killer 的典型特征（SIGKILL，不带错误码）。

解决方案建议：

• 把模型换成更小的量化版本（如 Q5_K_M → Q4_K_M，或者换小模型如 SDXL Turbo 的小尺寸版本）
• 减少图像分辨率（640x360 本身不大，没问题）
• 增加系统 swap 空间（临时代救）
• 如果有独显，优先用独显跑而非核显

简单说：核显显存不够，被系统 OOM 杀掉了，amd有共享显存设置，手动设置试试。以上为ai结论。