❓ Questions and Help
I am trying to run SenseVoice GGUF inference through the llama.cpp/ggml Vulkan backend on an Android device with a Mali-G615 MC6 GPU.
The Vulkan loader can enumerate the physical GPU successfully, but the process crashes with SIGSEGV during Vulkan backend initialization or the first compute pipeline creation,exit code 139. CPU inference works correctly on the same device, model, audio files, and runtime.
Is Mali-G615 / Mali r44p1 currently supported by the ggml Vulkan backend? Are there known limitations or required build flags for Android Mali GPUs, especially for integrated GPUs using unified memory?
The failure occurs with both Android NDK r28b and r27c builds.
Code
The SenseVoice CLI is executed with:
./llama-funasr-sensevoice \
--backend vulkan \
-m ./models/sensevoice-small-q8.gguf \
--vad ./models/fsmn-vad.gguf \
-a ./audio/00_BAC009S0770W0259.wav
❓ Questions and Help
I am trying to run SenseVoice GGUF inference through the llama.cpp/ggml Vulkan backend on an Android device with a Mali-G615 MC6 GPU.
The Vulkan loader can enumerate the physical GPU successfully, but the process crashes with
SIGSEGVduring Vulkan backend initialization or the first compute pipeline creation,exit code 139. CPU inference works correctly on the same device, model, audio files, and runtime.Is Mali-G615 / Mali r44p1 currently supported by the ggml Vulkan backend? Are there known limitations or required build flags for Android Mali GPUs, especially for integrated GPUs using unified memory?
The failure occurs with both Android NDK r28b and r27c builds.
Code
The SenseVoice CLI is executed with: