CUDA ARCH 9.0 unsupported?

[CHNtentes]

(qnn) root@gpu_h20:/nvme/shanghai-2/ltg/aimet/build# cmake .. -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DENABLE_CUDA=ON -DENABLE_TORCH=ON -DENABLE_TENSORFLOW=ON -DENABLE_ONNX=ON
-- The C compiler identification is GNU 9.4.0
-- The CXX compiler identification is GNU 9.4.0
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /usr/bin/cc - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Set SW_VERSION = 1.32.0 from /nvme/shanghai-2/ltg/aimet/packaging/version.txt
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD - Success
-- Found Threads: TRUE
Compiling with CUDA enabled
-- Found CUDAToolkit: /usr/local/cuda/targets/x86_64-linux/include (found version "11.8.89")
-- ** Initial CMAKE_CUDA_ARCHITECTURES = 52;60;61;70;72 **
-- The CUDA compiler identification is NVIDIA 11.8.89
-- Detecting CUDA compiler ABI info
-- Detecting CUDA compiler ABI info - done
-- Check for working CUDA compiler: /usr/local/cuda/bin/nvcc - skipped
-- Detecting CUDA compile features
-- Detecting CUDA compile features - done
-- Found CUDA toolkit version 11.8.89, using cu118
-- Found Python3: /root/miniconda3/envs/qnn/bin/python3.10 (found version "3.10.0") found components: Interpreter Development Development.Module Development.Embed
Found python: TRUE, at /root/miniconda3/envs/qnn/lib/libpython3.10.so
-- Found PkgConfig: /usr/bin/pkg-config (found version "0.29.1")
-- Checking for one of the modules 'lapacke'
-- Checking for one of the modules 'opencv'
-- Found ONNX version: 1.14.1
-- AIMET - ENABLE_TORCH: ON
-- Found Torch version: 2.1.2+cu118
-- ** Updated CMAKE_CUDA_ARCHITECTURES to 50;60;70;75;80;86;37;90 **
-- ** Removed unsupported archs (90), Now CMAKE_CUDA_ARCHITECTURES = 50;60;70;75;80;86;37 **
-- ** Updated TORCH_CUDA_ARCH_LIST to 5.0 6.0 7.0 7.5 8.0 8.6 3.7 **

[CHNtentes]

In CMakeLists.txt:

We remove certain architectures that are not supported

set(UNSUPPORTED_CUDA_ARCHITECTURES_TORCH 90)
list(REMOVE_ITEM CMAKE_CUDA_ARCHITECTURES ${UNSUPPORTED_CUDA_ARCHITECTURES_TORCH})
message(STATUS "** Removed unsupported archs (${UNSUPPORTED_CUDA_ARCHITECTURES_TORCH}),
Now CMAKE_CUDA_ARCHITECTURES = ${CMAKE_CUDA_ARCHITECTURES} **")

[CHNtentes]

@quic-akhobare @quic-bharathr @quic-mangal

[CHNtentes]

:(

[quic-akhobare]

Hi @CHNtentes - we are working to add this back. Some of the executors failed with arch 90 and was causing build instability. PR should be up soon.

[CHNtentes]

Hi @CHNtentes - we are working to add this back. Some of the executors failed with arch 90 and was causing build instability. PR should be up soon.

Thanks for your reply. Do I need to manually clone & build after PR is up, or you will release a 1.33.1 version?

[quic-jpolizzi]

@CHNtentes We will go ahead and provide a release once we get this completed, thanks!

[CHNtentes]

Hi, I saw that you merged a PR about this issue. Could you share when you will release a fix version? Our H20 server has not been in use for long time.