NPU models crash on Snapdragon 8 Elite (SM8750)

[Slayerx96]

Describe the bug:
All NPU models crash on loading.

To Reproduce:
Steps to reproduce the behavior:

1. Open the app
2. Tap on chat
3. Tap on Open on a NPU model tab [Gemma-4-E2B-it (Snapdragon 8 Elite) or Gemma 3 1B-IT (Snapdragon) or Gemma 3 270M (Snapdragon)]
4. Crash

Expected behavior:
The model correctly loads on NPU

Logcat:

box3.0.0_NPU-crash.txt

Device & App Information (Please complete the following):

• Device: Redmagic 11 Air
• Android Version: 16
• App Version: 3.0.0

Additional context:
Same Sigabrt crash for other gemma models, while trying to load Nano I get a crash regarding the MlKitContext initialization instead:

box3.0.0_Nano-crash.txt

[jegly]

Thanks for the detailed report and both logs — that's exactly what I needed to triage this , Open a new issue and reference issue #88 when you have the logs.

Nano crash: that one's expected, not a bug on your end. The Gemini Nano hub needs Google's on-device AICore, which only ships on Pixel 8+/select devices — your Red Magic doesn't have it, so MlKitContext is never initialized and the hub
throws. It won't affect anything else. I'll harden the hub so it shows an "unavailable on this device" message instead of crashing, but Nano models won't run on the SM8750 regardless. The NPU .litertlm models below are the ones meant
for your chip.

NPU crash (the real issue): this is a native abort deep inside the LiteRT runtime (Engine.initialize → "No usable Dispatch runtime found"), so the app can't catch it to fall back. v3.0.0 added a fix that makes the app load the Qualcomm
NPU driver instead of the Google Tensor one on Snapdragon devices, but your tombstone alone can't tell me whether that fix actually engaged on your device, or whether the Qualcomm driver loaded but couldn't run the model.
To pin it down, could you grab a full logcat (not just the crash tombstone) while reproducing? Plug in over USB and run

adb logcat -c
now reproduce: open chat → pick the Snapdragon NPU model → let it crash
adb logcat -d > box_npu_logcat.txt

Then attach box_npu_logcat.txt. The lines I'm after are anything mentioning litert_dispatch, dispatch_api, Qnn, GoogleTensor, or AGLlmChatModelHelper in the seconds before the crash — that tells me which driver loaded and why it bailed.

If adb isn't handy, a logcat-capture app (e.g. "Logcat Reader") filtered on the com.box.gallery process during the crash works too.

One more thing that helps: which model exactly did you pick — "Gemma-4-E2B-it (Snapdragon 8 Elite)", "Gemma 3 1B-IT (Snapdragon)", or "Gemma 3 270M (Snapdragon)" — and does it crash on all three?

[Slayerx96]

Using your test_build version I tried all three Gemma NPU models, the crash is the same, qnn_manager requires Qnn System library version 1.11.0 and find version 1.6.0 instead so DispatchDelegate fails to initialize.

box_npu_logcat.txt

[jegly]

sorry that was the old test build here is the confirmed working one https://drive.google.com/file/d/1PLU2PI8xzg8RdXJjpq0RMi2bCoCZP8Co/view?usp=sharing

a proper release should be out later today still polishing more bugs out