[Bug] CUDA kernel not found in registries for Op type: MMCVModulatedDeformConv2d

[chetir]

Checklist

• I have searched related issues but cannot get the expected help.
• 2. I have read the FAQ documentation but cannot get the expected help.
• 3. The bug has not been fixed in the latest version.

Describe the bug

I use mmcv.ops in my model. When run in pytorch, all works fine. But when export to onnx, it runs extremely slow, and i found dcn module may run on cpu. So how can I run dcn module on onnx cuda executor?

• torch exec time: 0.5s
• onnx exec time: 16s

Reproduction

export script example:

ort_custom_op_path = (
    "mmdeploy/mmdeploy/lib/libmmdeploy_onnxruntime_ops.so"
)
assert os.path.exists(ort_custom_op_path)
session_options = ort.SessionOptions()
session_options.register_custom_ops_library(ort_custom_op_path)


def export_onnx(model, model_path):
    input_0 = torch.zeros(1, 3, 1024, 1024).to(device)

    filename = model_path.replace("pth", "onnx")
    torch.onnx.export(
        model=model,
        args=(input_0,),
        f=filename,
        input_names=["input0"],
        output_names=["output0"],
        opset_version=11,
    )
    print("Finished onnx export")

Environment

04/30 18:11:42 - mmengine - INFO -

04/30 18:11:42 - mmengine - INFO - **********Environmental information**********
04/30 18:11:44 - mmengine - INFO - sys.platform: linux
04/30 18:11:44 - mmengine - INFO - Python: 3.7.7 (default, Mar 26 2020, 15:48:22) [GCC 7.3.0]
04/30 18:11:44 - mmengine - INFO - CUDA available: True
04/30 18:11:44 - mmengine - INFO - MUSA available: False
04/30 18:11:44 - mmengine - INFO - numpy_random_seed: 2147483648
04/30 18:11:44 - mmengine - INFO - GPU 0: NVIDIA GeForce RTX 3090
04/30 18:11:44 - mmengine - INFO - CUDA_HOME: /data/ao.xu/miniconda3
04/30 18:11:44 - mmengine - INFO - NVCC: Cuda compilation tools, release 11.3, V11.3.122
04/30 18:11:44 - mmengine - INFO - GCC: x86_64-conda_cos6-linux-gnu-gcc (crosstool-NG 1.23.0.449-a04d0) 7.3.0
04/30 18:11:44 - mmengine - INFO - PyTorch: 1.13.1+cu117
04/30 18:11:44 - mmengine - INFO - PyTorch compiling details: PyTorch built with:
  - GCC 9.3
  - C++ Version: 201402
  - Intel(R) oneAPI Math Kernel Library Version 2021.4-Product Build 20210904 for Intel(R) 64 architecture applications
  - Intel(R) MKL-DNN v2.6.0 (Git Hash 52b5f107dd9cf10910aaa19cb47f3abf9b349815)
  - OpenMP 201511 (a.k.a. OpenMP 4.5)
  - LAPACK is enabled (usually provided by MKL)
  - NNPACK is enabled
  - CPU capability usage: AVX2
  - CUDA Runtime 11.7
  - NVCC architecture flags: -gencode;arch=compute_37,code=sm_37;-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_80,code=sm_80;-gencode;arch=compute_86,code=sm_86
  - CuDNN 8.5
  - Magma 2.6.1
  - Build settings: BLAS_INFO=mkl, BUILD_TYPE=Release, CUDA_VERSION=11.7, CUDNN_VERSION=8.5.0, CXX_COMPILER=/opt/rh/devtoolset-9/root/usr/bin/c++, CXX_FLAGS= -fabi-version=11 -Wno-deprecated -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -fopenmp -DNDEBUG -DUSE_KINETO -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DSYMBOLICATE_MOBILE_DEBUG_HANDLE -DEDGE_PROFILER_USE_KINETO -O2 -fPIC -Wno-narrowing -Wall -Wextra -Werror=return-type -Werror=non-virtual-dtor -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wunused-local-typedefs -Wno-unused-parameter -Wno-unused-function -Wno-unused-result -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-stringop-overflow -Wno-psabi -Wno-error=pedantic -Wno-error=redundant-decls -Wno-error=old-style-cast -fdiagnostics-color=always -faligned-new -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Werror=cast-function-type -Wno-stringop-overflow, LAPACK_INFO=mkl, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, TORCH_VERSION=1.13.1, USE_CUDA=ON, USE_CUDNN=ON, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=ON, USE_NNPACK=ON, USE_OPENMP=ON, USE_ROCM=OFF,

04/30 18:11:44 - mmengine - INFO - TorchVision: 0.11.0
04/30 18:11:44 - mmengine - INFO - OpenCV: 4.11.0
04/30 18:11:44 - mmengine - INFO - MMEngine: 0.10.7
04/30 18:11:44 - mmengine - INFO - MMCV: 2.2.0
04/30 18:11:44 - mmengine - INFO - MMCV Compiler: GCC 9.3
04/30 18:11:44 - mmengine - INFO - MMCV CUDA Compiler: 11.7
04/30 18:11:44 - mmengine - INFO - MMDeploy: 1.3.1+3f8604b
04/30 18:11:44 - mmengine - INFO -

04/30 18:11:44 - mmengine - INFO - **********Backend information**********
04/30 18:11:44 - mmengine - INFO - tensorrt:    None
04/30 18:11:44 - mmengine - INFO - ONNXRuntime: None
04/30 18:11:44 - mmengine - INFO - ONNXRuntime-gpu:     1.14.1
04/30 18:11:44 - mmengine - INFO - ONNXRuntime custom ops:      Available
04/30 18:11:44 - mmengine - INFO - pplnn:       None
04/30 18:11:44 - mmengine - INFO - ncnn:        None
04/30 18:11:44 - mmengine - INFO - snpe:        None
04/30 18:11:44 - mmengine - INFO - openvino:    None
04/30 18:11:44 - mmengine - INFO - torchscript: 1.13.1
04/30 18:11:44 - mmengine - INFO - torchscript custom ops:      NotAvailable
04/30 18:11:44 - mmengine - INFO - rknn-toolkit:        None
04/30 18:11:44 - mmengine - INFO - rknn-toolkit2:       None
04/30 18:11:44 - mmengine - INFO - ascend:      None
04/30 18:11:44 - mmengine - INFO - coreml:      None
04/30 18:11:44 - mmengine - INFO - tvm: None
04/30 18:11:44 - mmengine - INFO - vacc:        None
04/30 18:11:44 - mmengine - INFO -

04/30 18:11:44 - mmengine - INFO - **********Codebase information**********
04/30 18:11:44 - mmengine - INFO - mmdet:       None
04/30 18:11:44 - mmengine - INFO - mmseg:       None
04/30 18:11:44 - mmengine - INFO - mmpretrain:  None
04/30 18:11:44 - mmengine - INFO - mmocr:       None
04/30 18:11:44 - mmengine - INFO - mmagic:      None
04/30 18:11:44 - mmengine - INFO - mmdet3d:     None
04/30 18:11:44 - mmengine - INFO - mmpose:      None
04/30 18:11:44 - mmengine - INFO - mmrotate:    None
04/30 18:11:44 - mmengine - INFO - mmaction:    None
04/30 18:11:44 - mmengine - INFO - mmrazor:     None
04/30 18:11:44 - mmengine - INFO - mmyolo:      None

Error traceback

2025-04-30 18:01:45.270207582 [I:onnxruntime:Default, cuda_execution_provider.cc:2382 GetCapability] CUDA kernel not found in registries for Op type: MMCVModulatedDeformConv2d node name: /dla_up/ida_0/proj_1/conv/MMCVModulatedDeformConv2d