Not Work on SOC SM7450 Snapdragon 7 Gen 1

[matrixportalx]

After many open source applications switched to Llama.rm, Snapdragon 7 Gen 1 SOC SM7450 is no longer working on my phone. I'd appreciate it if you could provide a solution for SM7450 SOC. If at least the NPU has an option for now, models other than Snapdragon 8 will also be able to continue to use open-weight models. Please do not restrict artificial intelligence only with Snapdragon 8 chipset devices. Your app, ChatterUI and PocketPal apps are no longer working on my Honor 90 phone. I would love to have Hexagon NPU technology working on my phone, but at least we can still use local artificial intelligence in the open source world, with at least CPU and CPU options.

Phone details:

Phone Model Name: Honor 90
Processor: Snapdragon 7 Gen 1 Accelerated Edition
RAM: 12 GB + 7.0 GB HONOR RAM Turbo
Soc: SM7450
Chipset: min
Device: HONOR REA-NX9
Software Version: MagicOS 9.0,
Android Version: 15

Bug details:
The application closes at the initialized stage of the model.

[jhen0409]

As readme mentioned, we currently only supported/tested 8 gen 1 or newer:

https://github.com/mybigday/llama.rn?tab=readme-ov-file#hexagon-npu-acceleration-experimental

The 7 gen 1 should choose other lib variant if available (rnllama_jni_v8*), if you can provide more details (like crash logs), I'll help to check.

[matrixportalx]

As readme mentioned, we currently only supported/tested 8 gen 1 or newer:

https://github.com/mybigday/llama.rn?tab=readme-ov-file#hexagon-npu-acceleration-experimental

The 7 gen 1 should choose other lib variant if available (rnllama_jni_v8*), if you can provide more details (like crash logs), I'll help to check.

Unfortunately, I don't have the chance to access the pc required for crash log for now

[matrixportalx]

In the Layla app, I can use it with the NPU engine by choosing HTP0' on my phone. It's much more vas than CPU, but it's working. I added the Layla log file with the details. I hope it helps.

llama_init.log

@ch_26f869e6-0819-4fe7-afee-a01c9a138256.log

[Vali-98]

They also submitted logs for ChatterUI:

Vali-98/ChatterUI#508

Here I found a very interesting line:

librnllama_v8_2_dotprod_i8mm_hexagon_opencl.so

The device somehow managed to load the i8mm_hexagon_opencl native library despite being Snapdragon 7 Gen 1. Iirc, ChatterUI does not alter how llama.rn loads native libs either.

@jhen0409 its possible that changes to the loadNative routine made it too permissive, causing this crash.

Their report also mentions ChatterUI v0.8.8 working, which was before llama.rn migrated to JSI.