RuntimeError: CUDA error: no kernel image is available for execution on the device

[gunshi]

Hi, I'm using detectron2 on a computing cluster and thus have various gpus that the code will be run on as per allocation. detectron was installed successfully and i'm able to import it from python.

However I get the following error on certain(most) gpus:

RuntimeError: CUDA error: no kernel image is available for execution on the device (ROIAlign_forward_cuda at /network/home/guptagun/od/detectron2_repo/detectron2/layers/csrc/ROIAlign/ROIAlign_cuda.cu:361)
frame #0: c10::Error::Error(c10::SourceLocation, std::string const&) + 0x47 (0x7f2459bbc687 in /network/home/guptagun/anaconda3/envs/detectron/lib/python3.7/site-packages/torch/lib/libc10.so)
frame #1: detectron2::ROIAlign_forward_cuda(at::Tensor const&, at::Tensor const&, float, int, int, int, bool) + 0xa24 (0x7f23f419189c in /network/home/guptagun/od/detectron2_repo/detectron2/_C.cpython-37m-x86_64-linux-gnu.so)
frame #2: detectron2::ROIAlign_forward(at::Tensor const&, at::Tensor const&, float, int, int, int, bool) + 0xb6 (0x7f23f4132f66 in /network/home/guptagun/od/detectron2_repo/detectron2/_C.cpython-37m-x86_64-linux-gnu.so)
frame #3: <unknown function> + 0x4ec8f (0x7f23f4144c8f in /network/home/guptagun/od/detectron2_repo/detectron2/_C.cpython-37m-x86_64-linux-gnu.so)
frame #4: <unknown function> + 0x49750 (0x7f23f413f750 in /network/home/guptagun/od/detectron2_repo/detectron2/_C.cpython-37m-x86_64-linux-gnu.so)
<omitting python frames>
frame #9: THPFunction_apply(_object*, _object*) + 0x8d6 (0x7f245a4abe96 in /network/home/guptagun/anaconda3/envs/detectron/lib/python3.7/site-packages/torch/lib/libtorch_python.so)

While on some gpus (one of them being Geforce GTX) the code runs as expected.

I was trying to run the demo.py file through:

python detectron2_repo/demo/demo.py --config-file detectron2_repo/configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml --input ./leftImg8bit.png --opts MODEL.WEIGHTS detectron2://COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x/137849600/model_final_f10217.pkl

Environment

output of python -m detectron2.utils.collect_env.

------------------------  --------------------------------------------------
sys.platform              linux
Python                    3.7.5 (default, Oct 25 2019, 15:51:11) [GCC 7.3.0]
Numpy                     1.15.4
Detectron2 Compiler       GCC 7.4
Detectron2 CUDA Compiler  10.0
DETECTRON2_ENV_MODULE     <not set>
PyTorch                   1.3.0
PyTorch Debug Build       False
torchvision               0.4.1a0+d94043a
CUDA available            True
GPU 0                     GeForce GTX TITAN X
CUDA_HOME                 None
Pillow                    5.3.0
cv2                       4.1.0
------------------------  --------------------------------------------------
PyTorch built with:
  - GCC 7.3
  - Intel(R) Math Kernel Library Version 2019.0.4 Product Build 20190411 for Intel(R) 64 architecture applications
  - Intel(R) MKL-DNN v0.20.5 (Git Hash 0125f28c61c1f822fd48570b4c1066f96fcb9b2e)
  - OpenMP 201511 (a.k.a. OpenMP 4.5)
  - NNPACK is enabled
  - CUDA Runtime 10.1
  - NVCC architecture flags: -gencode;arch=compute_35,code=sm_35;-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_61,code=sm_61;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_50,code=compute_50
  - CuDNN 7.6.3
  - Magma 2.5.1
  - Build settings: BLAS=MKL, BUILD_NAMEDTENSOR=OFF, BUILD_TYPE=Release, CXX_FLAGS= -Wno-deprecated -fvisibility-inlines-hidden -fopenmp -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -O2 -fPIC -Wno-narrowing -Wall -Wextra -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-sign-compare -Wno-unused-parameter -Wno-unused-variable -Wno-unused-function -Wno-unused-result -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-stringop-overflow -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 -Wno-stringop-overflow, DISABLE_NUMA=1, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, USE_CUDA=True, 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_STATIC_DISPATCH=OFF, 

when i build detectron using : python setup.py build develop
TORCH_CUDA_ARCH_LIST was set empty, and so it should have been compiled for all architectures? (acc to #62 (comment))
What can I do while compiling so that I'm able to use detectron on most gpus, or is this an issue with the compute node I'm using?

Thanks,
Gunshi

[ppwwyyxx]

Detectron2 CUDA Compiler 10.0

• CUDA Runtime 10.1

Your cuda versions mismatch and that's not allowed.

[gunshi]

Hi, I have actually tried loading both cuda/10.0 and cuda/10.1 modules one by one before running the demo.py command and i still get the error, are you saying i should install pytorch and then detectron2 again but after loading cuda 10.0 specifically?

[ppwwyyxx]

You should in general look at python -m detectron2.utils.collect_env to see whether this has been fixed prior to running detectron2.
You need to install the pytorch that matches your cuda module and recompile detectron2 afterwards.

[dhaivat1729]

@ppwwyyxx I am facing the exact same issue and my pytorch and detectron2 are compiled with exact same cuda versions. Also, I am facing this issue when I try to run detectron2 on different GPU than the one I have used to compile it. Here I compiled with titanX GPU, so it doesn't work on titanrtx or other GPUs. Note that I haven't installed using pip as I am modifying the codebase(only python files, not touching any cuda implementation) for my research, not sure if that has any effect though. Here is the output of python -m detectron2.utils.collect_env. Could you try installing on one GPU and test on other and see if this is general issue or I messed something up.

------------------------  --------------------------------------------------
sys.platform              linux
Python                    3.7.4 (default, Aug 13 2019, 20:35:49) [GCC 7.3.0]
Numpy                     1.17.2
Detectron2 Compiler       GCC 7.4
Detectron2 CUDA Compiler  10.0
DETECTRON2_ENV_MODULE     <not set>
PyTorch                   1.3.0
PyTorch Debug Build       False
torchvision               0.4.1a0+d94043a
CUDA available            True
GPU 0                     TITAN V
CUDA_HOME                 /ai/apps/cuda/10.0
NVCC                      Cuda compilation tools, release 10.0, V10.0.130
Pillow                    6.2.0
cv2                       4.1.0
------------------------  --------------------------------------------------
PyTorch built with:
  - GCC 7.3
  - Intel(R) Math Kernel Library Version 2019.0.4 Product Build 20190411 for Intel(R) 64 architecture applications
  - Intel(R) MKL-DNN v0.20.5 (Git Hash 0125f28c61c1f822fd48570b4c1066f96fcb9b2e)
  - OpenMP 201511 (a.k.a. OpenMP 4.5)
  - NNPACK is enabled
  - CUDA Runtime 10.0
  - NVCC architecture flags: -gencode;arch=compute_35,code=sm_35;-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_61,code=sm_61;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_50,code=compute_50
  - CuDNN 7.6.3
  - Magma 2.5.1
  - Build settings: BLAS=MKL, BUILD_NAMEDTENSOR=OFF, BUILD_TYPE=Release, CXX_FLAGS= -Wno-deprecated -fvisibility-inlines-hidden -fopenmp -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -O2 -fPIC -Wno-narrowing -Wall -Wextra -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-sign-compare -Wno-unused-parameter -Wno-unused-variable -Wno-unused-function -Wno-unused-result -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-stringop-overflow -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 -Wno-stringop-overflow, DISABLE_NUMA=1, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, USE_CUDA=True, 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_STATIC_DISPATCH=OFF, 

[ppwwyyxx]

I am facing this issue when I try to run detectron2 on different GPU than the one I have used to compile it.

answered in https://github.com/facebookresearch/detectron2/blob/master/INSTALL.md#common-installation-issues

[dhaivat1729]

@ppwwyyxx can you tell me if there is any way to set TORCH_CUDA_ARCH_LIST such that it works across all GPUs which support a particular cuda version? I see this variable set in Dockerfile for a set of GPUs, is there any command to set it for all? Like TORCH_CUDA_ARCH_LIST=All, rather than specifying GPUs TORCH_CUDA_ARCH_LIST="Maxwell;Maxwell+Tegra;Pascal;Volta;Turing"

[ppwwyyxx]

This is a pytorch question and you can refer to pytorch/pytorch#18781
It does not seem like "All" works for extension compilation at the moment.

[davidRetana]

I have the same problem trying to execute the following code:

>>> import torch
>>> device = torch.device("cuda")
>>> torch.rand(10).to(device)

RuntimeError: CUDA error: no kernel image is available for execution on the device

I solved the problem changing the driver from open source to NVIDIA proprietary (Ubuntu 18.04)

before: Using NVIDIA driver metapackage from nvidia-driver-410 (open source)
after: Using NVIDIA driver metapackage from nvidia-driver-390 (proprietary)

>>> torch.rand(10).to(device)
tensor([0.9129, 0.8937, 0.7499, 0.5510, 0.5670, 0.9313, 0.3335, 0.4019, 0.2288,
        0.4771])

Hope it helps.

[veronikayurchuk]

Hi @ppwwyyxx! I am having the same problem and the solution from Common Installation Issues didn't help.
I am running the code on Tesla V100 and getting an error. Running on Tesla K80 didn't produce such an error.

The error:

RuntimeError: CUDA error: no kernel image is available for execution on the device (ROIAlign_forward_cuda at /home/veronica/detectron2/detectron2/layers/csrc/ROIAlign/ROIAlign_cuda.cu:364)
frame #0: c10::Error::Error(c10::SourceLocation, std::string const&) + 0x33 (0x7fadadba1193 in /home/veronica/dirs/detectron2/env2/lib/python3.7/site-packages/torch/lib/libc10.so)
frame #1: detectron2::ROIAlign_forward_cuda(at::Tensor const&, at::Tensor const&, float, int, int, int, bool) + 0x9f4 (0x7fada3f24f2d in /home/veronica/dirs/detectron2/detectron2/_C.cpython-37m-x86_64-linux-gnu.so)
frame #2: detectron2::ROIAlign_forward(at::Tensor const&, at::Tensor const&, float, int, int, int, bool) + 0x9c (0x7fada3ea436c in /home/veronica/dirs/detectron2/detectron2/_C.cpython-37m-x86_64-linux-gnu.so)
frame #3: + 0x55465 (0x7fada3eb5465 in /home/veronica/dirs/detectron2/detectron2/_C.cpython-37m-x86_64-linux-gnu.so)
frame #4: + 0x555fe (0x7fada3eb55fe in /home/veronica/dirs/detectron2/detectron2/_C.cpython-37m-x86_64-linux-gnu.so)
frame #5: + 0x4fe33 (0x7fada3eafe33 in /home/veronica/dirs/detectron2/detectron2/_C.cpython-37m-x86_64-linux-gnu.so)

Output of python -m detectron2.utils.collect_env:

`------------------------ ----------------------------------------------------------------------------------
sys.platform linux
Python 3.7.4 (default, Aug 13 2019, 20:35:49) [GCC 7.3.0]
numpy 1.18.1
detectron2 0.1 @/home/veronica/dirs/detectron2/detectron2
detectron2 compiler GCC 6.3
detectron2 CUDA compiler 10.1
detectron2 arch flags sm_37
DETECTRON2_ENV_MODULE
PyTorch 1.4.0 @/home/veronica/dirs/detectron2/env2/lib/python3.7/site-packages/torch
PyTorch debug build False
CUDA available True
GPU 0 Tesla V100-SXM2-16GB
CUDA_HOME /usr/local/cuda
NVCC Cuda compilation tools, release 10.1, V10.1.243
TORCH_CUDA_ARCH_LIST 6.0;6.1;6.2;7.0;7.5
Pillow 7.0.0
torchvision 0.5.0 @/home/veronica/dirs/detectron2/env2/lib/python3.7/site-packages/torchvision
torchvision arch flags sm_35, sm_50, sm_60, sm_70, sm_75
cv2 4.2.0

---

PyTorch built with:

• GCC 7.3
• Intel(R) Math Kernel Library Version 2019.0.4 Product Build 20190411 for Intel(R) 64 architecture applications
• Intel(R) MKL-DNN v0.21.1 (Git Hash 7d2fd500bc78936d1d648ca713b901012f470dbc)
• OpenMP 201511 (a.k.a. OpenMP 4.5)
• NNPACK is enabled
• CUDA Runtime 10.1
• 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_61,code=sm_61;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_37,code=compute_37
• CuDNN 7.6.3
• Magma 2.5.1

---

`

"Detectron2 CUDA Compiler", "CUDA_HOME", "PyTorch built with - CUDA" all contain cuda libraries of the same version.
I also tried to run export TORCH_CUDA_ARCH_LIST=6.0,7.0 & python train_net.py and it didn't help, I got the same error.

[ppwwyyxx]

You need to rebuild detectron2 with export TORCH_CUDA_ARCH_LIST=6.0;7.0.
Or build on the machine where you run detectron2.

[parth1497]

@ppwwyyxx
Hi, I am new to using Pytorch. I am facing the above error. My system specifications are given below:

OS : Ubuntu 16.04
CUDA version - 10.1
Device count - 1
Device name - GeForce GT 720
Device capability - (3,5)
Pytorch version - 1.4.0

Can someone guide me as to how to resolve this issue?

Thanks in advance.

[ppwwyyxx]

All information about such issues are given in https://github.com/facebookresearch/detectron2/blob/master/INSTALL.md#common-installation-issues