Skip to content

[Bug] NPU backend crashes on Snapdragon 8+ gen 1 #517

Description

@HelleBenjamin

Description

LLMs won't load on NPU backend, stuck on loading and after a while times out.

Steps to Reproduce

  1. Download 0.0.103-beta2 beta version and install it
  2. Download an llm model, for example qwen3.5-0.8b
  3. Set inference backend to NPU and reload the model

Expected Behavior

The model should reload and use NPU as the backend.

Actual Behavior

Just stuck on loading model, then times out. On adb side, it shows a crash report.

Screenshots / Screen Recordings

none

Environment

  • Platform: Android
  • OS Version: 16
  • Device: Motorola Thinkphone
  • App Version: v0.0.103-beta2
  • Model in use: qwen3.5-0.8B, gguf at q4_0 quant

Logs

From adb with command adb logcat | grep -iE "qnn|htp|adsprpc|hexagon|opencl|fastrpc|avc: denied"

07-10 20:07:25.633  2501  2510 V adsprpc : platform_qdi_driver.c:692:0x4f84:1: ############################### Process on cDSP CRASHED!!!!!!! ########################################
07-10 20:07:25.633  2501  2510 V adsprpc : platform_qdi_driver.c:692:0x4f84:1: --------------------- Crash Details are furnished below ------------------------------------
07-10 20:07:25.633  2501  2510 V adsprpc : platform_qdi_driver.c:692:0x4f84:1: Process "/frpc/f074dee0 mqt_v_native" crashed in thread "0x  e7a4f0:work" for unknown reason
07-10 20:07:25.633  2501  2510 V adsprpc : platform_qdi_driver.c:692:0x4f84:1: Crashed Shared Object "./libggml-htp-v69.so" load address : 0x20000000 
07-10 20:07:25.633  2501  2510 V adsprpc : platform_qdi_driver.c:692:0x4f84:1: fastrpc_shell_unsigned_3 load address : D00000  and size : 129D88 
07-10 20:07:25.633  2501  2510 V adsprpc : platform_qdi_driver.c:692:0x4f84:1: Fault PC   :    0x20004B7C 
07-10 20:07:25.633  2501  2510 V adsprpc : platform_qdi_driver.c:692:0x4f84:1: LR         :    0x2002A65C 
07-10 20:07:25.633  2501  2510 V adsprpc : platform_qdi_driver.c:692:0x4f84:1: SP         :    0xE61280 
07-10 20:07:25.633  2501  2510 V adsprpc : platform_qdi_driver.c:692:0x4f84:1: Bad VA     :    0xE50280 
07-10 20:07:25.633  2501  2510 V adsprpc : platform_qdi_driver.c:692:0x4f84:1: FP         :    0xE624C8 
07-10 20:07:25.633  2501  2510 V adsprpc : platform_qdi_driver.c:692:0x4f84:1: SSR        :    0x1970428 
07-10 20:07:25.633  2501  2510 V adsprpc : platform_qdi_driver.c:692:0x4f84:1: Error code :    0x628 
07-10 20:07:25.633  2501  2510 V adsprpc : platform_qdi_driver.c:692:0x4f84:1: Call trace : 
07-10 20:07:25.633  2501  2510 V adsprpc : platform_qdi_driver.c:692:0x4f84:1: [<2002A65C>] op_tri+0x784:     (./libggml-htp-v69.so) 
07-10 20:07:25.633  2501  2510 V adsprpc : platform_qdi_driver.c:692:0x4f84:1: [<20004B7C>] dma_queue_flush+0x1C:     (./libggml-htp-v69.so) 
07-10 20:07:25.633  2501  2510 V adsprpc : platform_qdi_driver.c:692:0x4f84:1: ----------------------------- End of Crash Report --------------------------------------------------
07-10 20:07:25.633  2501  2510 V adsprpc : platform_qdi_driver.c:692:0x4f84:1: Please refer to Hexagon SDK documentation "<HEXAGON_SDK_ROOT>/docs/tools/debug.html" for debugging the user PD exceptions.
07-10 20:07:25.693   396  5106 E ai.offgridmobile: vendor/qcom/proprietary/adsprpc/src/fastrpc_apps_user.c:1388: Error 0x4e: remote_handle64_invoke failed for handle 0xc6ea9ad0, method 6 on domain 3 (sc 0x6000100) (errno Operation not permitted)
07-10 20:07:25.693   396  5106 E ai.offgridmobile: vendor/qcom/proprietary/adsprpc/src/dspqueue/dspqueue_cpu.c:1740::error: 78: (nErr = dspqueue_rpc_wait_signal(dq->dsp_handle, &signal)) == 0
07-10 20:07:25.693   396  5106 E ai.offgridmobile: vendor/qcom/proprietary/adsprpc/src/dspqueue/dspqueue_cpu.c:1782: dspqueue_receive_signal_thread failed with 78 errno Operation not permitted
07-10 20:07:25.703   396  5106 E ai.offgridmobile: vendor/qcom/proprietary/adsprpc/src/fastrpc_apps_user.c:1388: Error 0x8000040d: remote_handle64_invoke failed for handle 0xc6ec9f10, method 3 on domain 3 (sc 0x3000000) (errno Success)
07-10 20:07:25.703   396  5106 E ai.offgridmobile: vendor/qcom/proprietary/adsprpc/src/fastrpc_apps_user.c:3516: exit_thread current process1 thread exit nErr = 0x8000040d domain 3, handle 0xb4000077c6ec9f10
07-10 20:07:25.714   396  5102 E ai.offgridmobile: vendor/qcom/proprietary/adsprpc/src/fastrpc_apps_user.c:1366: Error 0x4e: remote_handle_invoke failed for handle 0x3, method 4 on domain 3 (sc 0x4020200) (errno Operation not permitted)
07-10 20:07:25.723   396  5102 E ai.offgridmobile: vendor/qcom/proprietary/adsprpc/src/fastrpc_apps_user.c:1366: Error 0x8000040d: remote_handle_invoke failed for handle 0x3, method 4 on domain 3 (sc 0x4020200) (errno Success)
07-10 20:07:25.724   396  5102 E ai.offgridmobile: vendor/qcom/proprietary/adsprpc/src/listener_android.c:226:Error 0x8000040d: listener response with result 0x0 for ctx 0x6383, handle 0x45eaf00, sc 0xffffffff failed : listener thread exited (errno Success)
07-10 20:07:25.733   396  5102 E ai.offgridmobile: vendor/qcom/proprietary/adsprpc/src/fastrpc_apps_user.c:1388: Error 0x8000040d: remote_handle64_invoke failed for handle 0xc6ec9f10, method 3 on domain 3 (sc 0x3000000) (errno Success)
07-10 20:07:25.734   396  5102 E ai.offgridmobile: vendor/qcom/proprietary/adsprpc/src/fastrpc_apps_user.c:3516: exit_thread current process1 thread exit nErr = 0x8000040d domain 3, handle 0xb4000077c6ec9f10

Additional Context

Models work in other backends.

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions