Compilation of <torch/extension.h> error on Windows CUDA 11.5

[atalman]

We have following error when compiling CUDA 11.5 on windows

C:\actions-runner_work\pytorch\pytorch\build\win_tmp\build\torch\include\pybind11\cast.h(1429): error: too few arguments for temp
late template parameter "Tuple"
detected during instantiation of class "pybind11::detail::tuple_caster<Tuple, Ts...> [with Tuple=std::pair, Ts=<T1, T2>]
"
(1507): here

C:\actions-runner_work\pytorch\pytorch\build\win_tmp\build\torch\include\pybind11\cast.h(1503): error: too few arguments for temp
late template parameter "Tuple"
detected during instantiation of class "pybind11::detail::tuple_caster<Tuple, Ts...> [with Tuple=std::pair, Ts=<T1, T2>]
"
(1507): here

Complete failure log:
https://github.com/pytorch/pytorch/runs/4408796098?check_suite_focus=true

This looks like the same issue as this one:
facebookresearch/pytorch3d#843

Here is the workaround for this issue:
facebookresearch/pytorch3d@cb170ac

cc @ezyang @gchanan @zou3519 @peterjc123 @mszhanyi @skyline75489 @nbcsm @brianjo @mruberry @ngimel @bdhirsh @jbschlosser @malfet @seemethere @pytorch/pytorch-dev-infra

[seemethere]

Adding this to the 1.11.0 milestone since this will be a blocker for the 1.11.0 release

[malfet]

Grabbing for myself, as it probably involves tweaking pybind a bit

[malfet]

I can reproduce it with hello world example that simply includes pybind, which looks like a CUDA compiler bug (happens during the invocation of cicc):

#include <stdio.h>
#include <pybind11/pybind11.h>
__global__ void kernel() {
  printf("Hello World");
}
int main(void) {
 kernel<<<1, 1>>>();
 return cudaDeviceSynchronize();
}

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.5\bin>nvcc "c:\Users\runneruser\Documents\hello.cu" -o c:\Users\runneruser\Documents\a.exe -IC:\actions-runner\_work\pytorch\pytorch\third_party\pybind11\include -IC:\Jenkins\Miniconda3\include
hello.cu
...
#$ cicc --microsoft_version=1928 --msvc_target_version=1928 --compiler_bindir "C:/Program Files (x86)/Microsoft Visual Studio/2019/BuildTools/VC/Tools/MSVC/14.28.29333/bin/Hostx64/x64/../../../../../../.." --sdk_dir "C:/Program Files (x86)/Windows Kits/10/" --display_error_number --orig_src_file_name "c:/Users/runneruser/Documents/hello.cu" --orig_src_path_name "c:\Users\runneruser\Documents\hello.cu" --allow_managed  -arch compute_52 -m64 --no-version-ident -ftz=0 -prec_div=1 -prec_sqrt=1 -fmad=1 --include_file_name "tmpxft_00001bb8_00000000-7_hello.fatbin.c" -tused --gen_module_id_file --module_id_file_name "C:/Users/RUNNER~1/AppData/Local/Temp/2/tmpxft_00001bb8_00000000-8_hello.module_id" --gen_c_file_name "C:/Users/RUNNER~1/AppData/Local/Temp/2/tmpxft_00001bb8_00000000-10_hello.cudafe1.c" --stub_file_name "C:/Users/RUNNER~1/AppData/Local/Temp/2/tmpxft_00001bb8_00000000-10_hello.cudafe1.stub.c" --gen_device_file_name "C:/Users/RUNNER~1/AppData/Local/Temp/2/tmpxft_00001bb8_00000000-10_hello.cudafe1.gpu"  "C:/Users/RUNNER~1/AppData/Local/Temp/2/tmpxft_00001bb8_00000000-13_hello.cpp1.ii" -o "C:/Users/RUNNER~1/AppData/Local/Temp/2/tmpxft_00001bb8_00000000-10_hello.ptx"
C:\actions-runner\_work\pytorch\pytorch\third_party\pybind11\include\pybind11\detail/common.h(810): warning #1388-D: base class dllexport/dllimport specification differs from that of the derived class

C:\actions-runner\_work\pytorch\pytorch\third_party\pybind11\include\pybind11\pytypes.h(338): warning #1388-D: base class dllexport/dllimport specification differs from that of the derived class

C:\actions-runner\_work\pytorch\pytorch\third_party\pybind11\include\pybind11\pytypes.h(387): warning #1394-D: field of class type without a DLL interface used in a class with a DLL interface

C:\actions-runner\_work\pytorch\pytorch\third_party\pybind11\include\pybind11\pytypes.h(387): warning #1394-D: field of class type without a DLL interface used in a class with a DLL interface

C:\actions-runner\_work\pytorch\pytorch\third_party\pybind11\include\pybind11\pytypes.h(387): warning #1394-D: field of class type without a DLL interface used in a class with a DLL interface

C:\actions-runner\_work\pytorch\pytorch\third_party\pybind11\include\pybind11\cast.h(567): error: too few arguments for template template parameter "Tuple"
          detected during instantiation of class "pybind11::detail::tuple_caster<Tuple, Ts...> [with Tuple=std::pair, Ts=<T1, T2>]"
(648): here

C:\actions-runner\_work\pytorch\pytorch\third_party\pybind11\include\pybind11\cast.h(644): error: too few arguments for template template parameter "Tuple"
          detected during instantiation of class "pybind11::detail::tuple_caster<Tuple, Ts...> [with Tuple=std::pair, Ts=<T1, T2>]"
(648): here

2 errors detected in the compilation of "c:/Users/runneruser/Documents/hello.cu".
# --error 0x1 --

For debugging: hello.cpp1.ii

[zasdfgbnm]

Further narrow down:

#include <utility>

// Base implementation for std::tuple and std::pair
template <template<typename...> class Tuple, typename... Ts> class tuple_caster {
    using type = Tuple<Ts...>;
};

template <typename T1, typename T2> class type_caster
    : public tuple_caster<std::pair, T1, T2> {};


__global__ void kernel() {
	printf("Hello World");
}
int main(void) {
	kernel <<<1, 1>>> ();
	return cudaDeviceSynchronize();
}

[malfet]

Need to document the regression of nvcc compiler from CUDA-11.5 on Windows

[mszhanyi]

@malfet @atalman
I'd created a project GPU and CUDA regression on Windows to collect GPU and cuda regressions.

[MinttHu]

My system is window 10+CUDA11.3. My problem is that when I use "torch. utils.cpp_extension.load_inline" to built a C++/CUDA extension, it shows: "ninja: build stopped: subcommand failed."

[3a1b2c3]

Any workarounds?

[atalman]

@malfet @ptrblck Since we have cuda 11.7 I will rerun testing on this cuda to see if this issue is resolved for CUDA 11.7

[atalman]

Looks like this issue is resolved in 11.7. I am not observing the failure anymore on #85966

[atalman]

Removing from milestones, since this issue is not critical for release 1.13 and this is resolved for CUDA 11.7

[atalman]

With deprecation of CUDA 11.6, we can resolve this issue, will post a PR once 11.6 is deprecated from CI