mlc-ai
[Bug] how to accurately measure the real memory usage on Android ？
After deploying Qwen-4B, Qwen-1.8B, and Qwen-0.5B to an Android phone using mlc-llm, I used the following commands to check the system memory usage:
adb shell ps | findstr "mlc"
adb shell dmabuf_dump PID
adb shell dumpsys meminfo PID
However, the results were all similar, showing only a few hundred MB, so I suspect there might be an issue with the method I'm using. Could you please advise on how to accurately measure the real memory usage when mlc-llm is deployed on a phone, both during model loading and runtime? For example, are there any specific commands you would recommend?
Android Studio may have some tools for memory profiling https://developer.android.com/studio/profile/memory-profiler.
a few hundred MB must be the CPU memory usage. However, the model is stored in the GPU memory, so OS-level memory command is not enough
a few hundred MB must be the CPU memory usage. However, the model is stored in the GPU memory, so OS-level memory command is not enough
@Hzfengsy Is there a more detailed command?
Android Studio may have some tools for memory profiling https://developer.android.com/studio/profile/memory-profiler.
@MasterJH5574 You can try, but this method also cannot accurately display memory usage, only a few hundred megabytes.
I used Android Studio tools for memory profiling, and got the result. It seems to be correct. The left and right peaks represent two models of different sizes
Yes I could also see accurate memory usage of gigabytes when I used it last time.
Almost perfect sollution:
watch -n0.1 'adb shell "cat /proc/meminfo | head"'
- read how memory works on android (e.g. https://developer.android.com/topic/performance/memory-management)
