🐛 Description
App crashes instantly upon clicking the "Generate Image" (创作图片) button on a Snapdragon 8 Elite device running version 5.6.0 (unpacked/unprotected APK).
According to the system crash log, the native library libsdloader.so throws an unhandled exception when initializing the computation graph and submitting tasks to the Vulkan queue, leading to a fatal vk::DeviceLostError.
Core Crash Log:
Abort message: 'terminating due to uncaught exception of type vk::DeviceLostError: vk::QueueSubmit: ErrorDeviceLost'
#00 pc 000000000007160c /apex/com.android.runtime/lib64/bionic/libc.so (abort+160)
#1 pc 00000000031013f4 /data/app/~~3yqi1rClY7UBYqPgUMz1sg==/org.onion.diffusion-DB4LvF9mqjf0UXOoadp4mg==/base.apk!libsdloader.so
...
#8 pc 000000000307cddc /data/app/.../base.apk!libsdloader.so (ggml_backend_graph_compute+24)
#9 pc 0000000002e3c640 /data/app/.../base.apk!libsdloader.so (GGMLRunner::compute(std::__ndk1::function<ggml_cgraph* ()>, int, bool, ggml_tensor**, ggml_context*)+664)
💻 System Information
OS Version: Android 15 / Xiaomi HyperOS (with 16KB Page Size kernel)
RAM/Memory: 16GB RAM + 1TB Storage
CPU: Snapdragon 8 Elite
Device Model: Xiaomi 15 Ultra (xuanyuan)
🕒 Issue Scenario
[ ] Model Loading: Error occurred while selecting or initializing the model.
[x] Image Generation: Error occurred during the inference process.
[ ] UI/UX: Visual glitches or interaction issues.
[ ] Other: (Please specify below)
⚙️ Settings Used
Model Path: majicmixRealistic_v7.safetensors
Quantization (wtype): F16 / Auto (Standard Safetensors format)
Offload to CPU: Disabled (Using GPU/Vulkan backend)
Enable MMAP: Enabled
Flash Attention: [Enabled/Disabled] (Please change based on your actual toggle)
🛠 Reproduction Steps
Open the app, put and load the majicmixRealistic_v7.safetensors model with normal generation parameters.
Click the "Generate Image" (创作图片) button.
The app initializes the compute graph, calls the Vulkan backend, and immediately crashes right at the moment of GGMLRunner::compute queue submission.
📸 Screenshots：
🐛 Description
App crashes instantly upon clicking the "Generate Image" (创作图片) button on a Snapdragon 8 Elite device running version 5.6.0 (unpacked/unprotected APK).
According to the system crash log, the native library libsdloader.so throws an unhandled exception when initializing the computation graph and submitting tasks to the Vulkan queue, leading to a fatal vk::DeviceLostError.
Core Crash Log:
Abort message: 'terminating due to uncaught exception of type vk::DeviceLostError: vk::QueueSubmit: ErrorDeviceLost'
#00 pc 000000000007160c /apex/com.android.runtime/lib64/bionic/libc.so (abort+160)
#1 pc 00000000031013f4 /data/app/~~3yqi1rClY7UBYqPgUMz1sg==/org.onion.diffusion-DB4LvF9mqjf0UXOoadp4mg==/base.apk!libsdloader.so
...
#8 pc 000000000307cddc /data/app/.../base.apk!libsdloader.so (ggml_backend_graph_compute+24)
#9 pc 0000000002e3c640 /data/app/.../base.apk!libsdloader.so (GGMLRunner::compute(std::__ndk1::function<ggml_cgraph* ()>, int, bool, ggml_tensor**, ggml_context*)+664)
💻 System Information
OS Version: Android 15 / Xiaomi HyperOS (with 16KB Page Size kernel)
RAM/Memory: 16GB RAM + 1TB Storage
CPU: Snapdragon 8 Elite
Device Model: Xiaomi 15 Ultra (xuanyuan)
🕒 Issue Scenario
[ ] Model Loading: Error occurred while selecting or initializing the model.
[x] Image Generation: Error occurred during the inference process.
[ ] UI/UX: Visual glitches or interaction issues.
[ ] Other: (Please specify below)
⚙️ Settings Used
Model Path: majicmixRealistic_v7.safetensors
Quantization (wtype): F16 / Auto (Standard Safetensors format)
Offload to CPU: Disabled (Using GPU/Vulkan backend)
Enable MMAP: Enabled
Flash Attention: [Enabled/Disabled] (Please change based on your actual toggle)
🛠 Reproduction Steps
Open the app, put and load the majicmixRealistic_v7.safetensors model with normal generation parameters.
Click the "Generate Image" (创作图片) button.
The app initializes the compute graph, calls the Vulkan backend, and immediately crashes right at the moment of GGMLRunner::compute queue submission.
📸 Screenshots：