make building stop at Built target portable_kernels

[imjking]

Hey, i want to export llama pte model and deploy it on SA8255 device, i refered to https://github.com/pytorch/executorch/blob/main/examples/models/llama/README.md and https://docs.pytorch.ac.cn/executorch/stable/llm/build-run-llama3-qualcomm-ai-engine-direct-backend.html, but when i Built llama runner binary for Android i got the error:
[ 98%] Linking CXX static library libportable_kernels.a
[ 98%] Built target portable_kernels
[ 98%] Linking CXX static library liboptimized_portable_kernels.a
[ 98%] Built target optimized_portable_kernels
gmake: *** [Makefile:156: all] Error 2

How can I solve this? i paste the error.log and the build.sh file.
Appreciate reply!

build.sh
error.log

cc @cccclai @winskuo-quic @shewu-quic @haowhsu-quic @DannyYuyang-quic @cbilgin

[GregoryComer]

Hi @imjking. It looks like it's not finding one of the submodules (pocketfft). Can you try running git submodule update --init from the executorch directory? After that, you should hopefully see pocketfft_hdronly.h in third-party/pocketfft.

[imjking]

@GregoryComer
Thanks, i sovled the problem pocketfft problem.
but when i run llama_main on 8255, i got Missing operator error as follow:

E 00:00:03.098432 executorch:method.cpp:788] Missing operator: [11] llama::custom_sdpa.out
E 00:00:03.098450 executorch:operator_registry.cpp:256] kernel 'llama::fast_hadamard_transform.out' not found.
I 00:00:03.098452 executorch:operator_registry.cpp:257] dtype: 6 | dim order: [
I 00:00:03.098453 executorch:operator_registry.cpp:257] 0,
I 00:00:03.098455 executorch:operator_registry.cpp:257] 1,
I 00:00:03.098456 executorch:operator_registry.cpp:257] 2,
I 00:00:03.098475 executorch:operator_registry.cpp:257] ]
I 00:00:03.098517 executorch:operator_registry.cpp:257] dtype: 6 | dim order: [
I 00:00:03.098519 executorch:operator_registry.cpp:257] 0,
I 00:00:03.098540 executorch:operator_registry.cpp:257] 1,
I 00:00:03.098555 executorch:operator_registry.cpp:257] 2,
I 00:00:03.098557 executorch:operator_registry.cpp:257] ]
I 00:00:03.098604 executorch:operator_registry.cpp:257] dtype: 6 | dim order: [
I 00:00:03.098607 executorch:operator_registry.cpp:257] 0,
I 00:00:03.098631 executorch:operator_registry.cpp:257] 1,
I 00:00:03.098632 executorch:operator_registry.cpp:257] 2,
I 00:00:03.098634 executorch:operator_registry.cpp:257] ]
E 00:00:03.098635 executorch:method.cpp:788] Missing operator: [12] llama::fast_hadamard_transform.out
E 00:00:03.098650 executorch:method.cpp:1053] There are 64 instructions don't have corresponding operator registered. See logs for details
E 00:00:03.152838 executorch:main.cpp:181] Failed to run llama runner

Is it the llama_main file is not compiled correctly? When i rebuilt build.sh file with following additional params:
-DEXECUTORCH_BUILD_XNNPACK=ON
-DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON
-DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON
-DEXECUTORCH_BUILD_KERNELS_LLM=ON \

i got a problem at the end like:
CMake Error at oss_scripts/llama/CMakeLists.txt:13 (add_library):
add_library cannot create target "custom_ops" because an imported target
with the same name already exists.
CMake Error at oss_scripts/llama/CMakeLists.txt:14 (target_include_directories):
target_include_directories may only set INTERFACE properties on IMPORTED
targets
CMake Error at oss_scripts/llama/CMakeLists.txt:15 (target_include_directories):
target_include_directories may only set INTERFACE properties on IMPORTED
targets
CMake Error at oss_scripts/llama/CMakeLists.txt:18 (target_link_libraries):
Cannot specify link libraries for target "custom_ops" which is not built by
this project.

How can i solve the Missing operator problem on 8255 device?

[cccclai]

Ah I think you're building the runner for cpu but the model you export is for qnn. Just a few checklist:

• Did you follow the guidance in https://docs.pytorch.org/executorch/1.0/backends-qualcomm.html to install the qnn backend?
• Are you able to run any simple model https://github.com/pytorch/executorch/tree/main/examples/qualcomm#simple-examples-to-verify-the-backend-is-working on SA8255?

Also, to double check, are you trying to run the models on CPU or NPU? The guidance above is to run it on NPU

[imjking]

@cccclai
Yes, I'm trying to run the llama3.2 1B or llama3.2 3B model on NPU(8255).

1. i have installed the qnn backend 2.37 on x86, ubuntu22.04
2. i will try to run the simple model to verify on SA8255
   Thanks.

[imjking]

@cccclai
Sorry to bother.
Now i use https://github.com/pytorch/executorch/tree/main/examples/qualcomm#simple-examples-to-verify-the-backend-is-working on SA8255 to verify, but when i run ’python export_example.py -m add -g‘, i got following errors, how to solve this?

Skipping import of cpp extensions due to incompatible torch version 2.9.0+cu128 for torchao version 0.14.0 Please see GitHub issue #2919 for more info
Traceback (most recent call last):
File "/workspace/npu_env/bin/flatc", line 3, in
from executorch.data.bin import flatc
ImportError: cannot import name 'flatc' from 'executorch.data.bin' (unknown location)
Traceback (most recent call last):
File "/workspace/qairt-2.37/qairt/executorch/examples/qualcomm/scripts/export_example.py", line 120, in
main() # pragma: no cover
File "/workspace/qairt-2.37/qairt/executorch/examples/qualcomm/scripts/export_example.py", line 102, in main
compile_spec = generate_qnn_executorch_compiler_spec(
File "/workspace/qairt-2.37/qairt/executorch/backends/qualcomm/utils/utils.py", line 1087, in generate_qnn_executorch_compiler_spec
CompileSpec(QCOM_QNN_COMPILE_SPEC, option_to_flatbuffer(qnn_executorch_options))
File "/workspace/qairt-2.37/qairt/executorch/backends/qualcomm/serialization/qc_schema_serialize.py", line 55, in option_to_flatbuffer
return _convert_to_flatbuffer(qnn_executorch_options, "qc_compiler_spec")
File "/workspace/qairt-2.37/qairt/executorch/backends/qualcomm/serialization/qc_schema_serialize.py", line 31, in _convert_to_flatbuffer
_flatc_compile(d, schema_path, json_path)
File "/workspace/qairt-2.37/qairt/executorch/exir/_serialize/_flatbuffer.py", line 213, in _flatc_compile
_run_flatc(
File "/workspace/qairt-2.37/qairt/executorch/exir/_serialize/_flatbuffer.py", line 199, in _run_flatc
subprocess.run([flatc_path] + list(args), check=True)
File "/usr/lib/python3.10/subprocess.py", line 526, in run
raise CalledProcessError(retcode, process.args,
subprocess.CalledProcessError: Command '['flatc', '--binary', '-o', '/tmp/tmpq32k1i8s', '/tmp/tmpq32k1i8s/qc_compiler_spec.fbs', '/tmp/tmpq32k1i8s/qc_compiler_spec.json']' returned non-zero exit status 1.

[cccclai]

How did you install executorch and qnn? these are the steps I usually do

Run these commands in executorch folder

1. install_executorch.sh
2. backends/qualcomm/scripts/build.sh # it needs to set up the env correctly

[imjking]

Ah I think you're building the runner for cpu but the model you export is for qnn. Just a few checklist:

• Did you follow the guidance in https://docs.pytorch.org/executorch/1.0/backends-qualcomm.html to install the qnn backend?
• Are you able to run any simple model https://github.com/pytorch/executorch/tree/main/examples/qualcomm#simple-examples-to-verify-the-backend-is-working on SA8255?

Also, to double check, are you trying to run the models on CPU or NPU? The guidance above is to run it on NPU

---

Hey, i tried https://github.com/pytorch/executorch/tree/main/examples/qualcomm#simple-examples-to-verify-the-backend-is-working to export add.pte and deeplab_v3.pte, and tested on 8255 device, and the result like below, can this prove that it is correct? :

Additionally， i still met the custom_op problem on 8255:

CMake Error at oss_scripts/llama/CMakeLists.txt:13 (add_library):
add_library cannot create target "custom_ops" because an imported target
with the same name already exists.
CMake Error at oss_scripts/llama/CMakeLists.txt:14 (target_include_directories):
target_include_directories may only set INTERFACE properties on IMPORTED
targets.

I referred to https://github.com/pytorch/executorch/tree/main/examples/models/qwen3 to export qwen3-1.7b.pte model and run executorch/backends/qualcomm/scripts/build.sh to build Android llama_main， am i right? How can i build the runner for npu and the model also for qnn?

[cccclai]

That's great! backends/qualcomm/scripts/build.sh should build all targets including the llama runner for you. Can you look for qnn_llama_runner (https://github.com/pytorch/executorch/blob/main/examples/qualcomm/oss_scripts/llama/CMakeLists.txt#L53) in the android build folder?

To generate the .pte model, can you try this instruction? https://github.com/pytorch/executorch/tree/main/examples/qualcomm/oss_scripts/llama#qwen3-17b Note that the calibration process will be long, please be patient (or use the smaller model like 0.6B to verify end to end first before spending too much time)

[imjking]

@cccclai Thanks so much for quick reply.

1. yes, i find qnn_llama_runner in the android build folder : ./examples/qualcomm/oss_scripts/llama/qnn_llama_runner. i tried qnn_llama_runner on 8255 but get another error:

Is it used in the same way as llama_main? Or how should I use it? Do I need to configure anything else or set up an environment on 8255 device?

2. OK，i will try following instruct to generate pte. https://github.com/pytorch/executorch/tree/main/examples/qualcomm/oss_scripts/llama#qwen3-17b.
   BTW, is this instruct generate diffierent *.pte compared with https://github.com/pytorch/executorch/tree/main/examples/models/qwen3 ‘s command like: python -m extension.llm.export.export_llm xxx ?

[cccclai]

From the error message, it seems like a linker error, what NDK version are you using?

[imjking]

I'm using the android-ndk-r27d version of NDK.

[cccclai]

@haowhsu-quic @shewu-quic @winskuo-quic @chenweng-quic any thoughts on this error?