A retail phone will not let you attach a GPU debugger to an arbitrary app, but it will load a Vulkan layer from your own APK if you package it and enable it from the app's own manifest. The layer writes a capture file that you copy off the device and open on a desktop.
Our workflow: build a debug variant with the capture layer packaged as a native library, set debug.vulkan.layers in the manifest, capture one frame from a reproducible camera position, then copy the capture with adb pull. The whole loop is under two minutes on a mid-range phone, which is short enough to use while tuning rather than only after a bug.
Two limits worth knowing. Captures of frames with more than about 4 000 draw calls get large fast; we capture a trimmed scene instead. And vendor drivers occasionally fail to capture a frame that uses sparse resources, which is a driver bug, not a reason to stop capturing.
Read next: Profiling one frame: where 16.6 ms goes.