App: Box v3.4.5 — Image Gen section
Device: Redmi Note 13 (Snapdragon 685, Android 15)
Model: LCM-SSD-1B Q4_K
When generating an image from the Image Gen section, generation runs through all steps but crashes right before the output image is produced. Native crash in libstablediffusion.so:
#00 ggml_vec_dot_q4_K_q8_K+136
#01 ggml_compute_forward_mul_mat+1964
#02 ggml_graph_compute_thread+1224
#03 __kmp_invoke_microtask+152 (libomp.so)
The crash happens inside the Q4_K quantized matmul dot-product kernel during multithreaded OpenMP compute, not in graph setup or I/O — looks like an instruction/SIMD path issue rather than a memory allocation problem. Given the SoC is a Snapdragon 685 (Cortex-A73/A55 cores), this may be the same class of ARM dotprod/FP16 kernel incompatibility seen before on mid-range Snapdragon chips — worth checking if the Q4_K dot-product kernel has a safe fallback for cores without the relevant SIMD extensions.
Happy to test a build with that kernel path patched/disabled if useful.
App: Box v3.4.5 — Image Gen section
Device: Redmi Note 13 (Snapdragon 685, Android 15)
Model: LCM-SSD-1B Q4_K
When generating an image from the Image Gen section, generation runs through all steps but crashes right before the output image is produced. Native crash in
libstablediffusion.so:The crash happens inside the Q4_K quantized matmul dot-product kernel during multithreaded OpenMP compute, not in graph setup or I/O — looks like an instruction/SIMD path issue rather than a memory allocation problem. Given the SoC is a Snapdragon 685 (Cortex-A73/A55 cores), this may be the same class of ARM dotprod/FP16 kernel incompatibility seen before on mid-range Snapdragon chips — worth checking if the Q4_K dot-product kernel has a safe fallback for cores without the relevant SIMD extensions.
Happy to test a build with that kernel path patched/disabled if useful.