無法將onnx模型編譯成nef格式
您好：
我在使用kneron toolchain且想將onnx 模型編譯成nef ，並且依照這個網頁內https://doc.kneron.com/docs/#toolchain/appendix/operators/ 查看支援的算子，但我已經排除了kl720不支援的算子，仍然有以下錯誤：
use toolchain binaries
use toolchain binaries
========================================
Processing model: best
INFO:kneronnxopt:The input model has been checked. Start simplifying.
Checking 0/1...
INFO:kneronnxopt:Start independent preprocessing optimization passes.
INFO:kneronnxopt:Start shape preparation and optimizations.
INFO:kneronnxopt:Start operator related optimizations.
INFO:kneronnxopt:Start pattern related optimizations.
INFO:kneronnxopt:Start independent postprocessing optimization passes.
INFO:kneronnxopt:Comparing inference result between original/optimized ONNX... (add flag '--skip_check' if want to skip this step)
WARNING:kneronnxopt.checker:
Arrays are not almost equal to 4 decimals
Mismatched elements: 99 / 210000 (0.0471%)
Max absolute difference: 0.00024414
Max relative difference: 5.7531443e-06
x: array([1.1304e+01, 1.7465e+01, 2.1873e+01, ..., 3.7253e-06, 2.3246e-06,
3.4571e-06], dtype=float32)
y: array([1.1304e+01, 1.7465e+01, 2.1873e+01, ..., 3.7253e-06, 2.3246e-06,
3.4571e-06], dtype=float32)
WARNING:kneronnxopt:The optimized model might not be correct.
INFO:kneronnxopt:The input model has been checked. Start simplifying.
Checking 0/1...
INFO:kneronnxopt:Start independent preprocessing optimization passes.
INFO:kneronnxopt:Start shape preparation and optimizations.
INFO:kneronnxopt:Start operator related optimizations.
INFO:kneronnxopt:Start pattern related optimizations.
INFO:kneronnxopt:Start independent postprocessing optimization passes.
INFO:kneronnxopt:Comparing inference result between original/optimized ONNX... (add flag '--skip_check' if want to skip this step)
INFO:kneronnxopt.checker:Two models have the same behaviour.
ERROR:root: Failure for model "best_optimized/best_optimized" when running "kdp720/compiler frontend"
===========================================
= report on flow status =
===========================================
kdp720 general
compiler frontend compiler_cfg clean_opt nef_model_id model oversize post_clean onnx size (MB)
category case
best_optimized best_optimized Err: 134 ✓ ✓ 32768 ✓ ✓ 9
[2025-11-25 15:24:10] [debug] [Thread: 834] [/projects_src/kneron-piano_v2/dynasty/floating_point/floating_point/src/common/BaseInferencerImpl.cpp:63]
start to create operators....
[2025-11-25 15:24:10] [debug] [Thread: 834] [/projects_src/kneron-piano_v2/dynasty/floating_point/floating_point/src/common/BaseInferencerImpl.cpp:102]
The model allocated space: 585.29252 Mbytes( including workspace: 0.0 Mbytes)
Section 3 E2E simulator finished.
ERROR:root: Failure for model "best_optimized/best_optimized" when running "kdp720/compiler frontend"
===========================================
= report on flow status =
===========================================
kdp720 general
compiler frontend compiler_cfg clean_opt nef_model_id model oversize post_clean onnx size (MB)
category case
best_optimized best_optimized Err: 134 ✓ ✓ 32768 ✓ ✓ 9
Workflow failed!
Quantization model generation failed. See above message for details.
在此附上我的模型以及編譯腳本
希望能幫我解惑，到底是哪裡有問題，導致無法編譯成功。
Comments
您好，
感謝您回報此問題。我們建議您將optimized後的模型裡面Concat以下的operators剪掉之後再去做轉換，如下圖。您可以再把剪掉的operators加回去postprocess裡面。
您好，
我想請問一下
我在嘗試將基於 YOLOv8 的目標檢測模型部署到 Kneron KL720
由於 KL720 硬體不支援 Softmax 算子，我將其替換為 4階泰勒展開 來近似
那我optimized後的模型裡面Concat以下的operators在參考https://doc.kneron.com/docs/#toolchain/appendix/operators/ 都有符合硬體算子的限制
也想要將模型全部都在npu上跑
想請問是哪裡出沒有符合限制才導致無法轉換
您好，
雖然模型的operators都有支援，但是onnx模型裡面的Slice導致Toolchain KL720的compiler無法轉換模型。我們嘗試剪掉operators後，就可以成功轉換了。因為KL720是一個比較舊的平台，會需要很多時間修改這些bugs，所以會建議您剪掉operators，或是更換模型，不好意思。
了解了
謝謝您
不好意思
我追問一下
所以是編譯器有bug嗎
還是我的onnx模型有問題，沒有符合編譯的限制??