Fails to run find_package(Torch) on Windows with libtorch package #333
Comments
|
Yeah, I just noticed this in conda-forge/torchaudio-feedstock#14. The issue is that we move In this way, this is a follow-up to of #327 & #328. I think the easiest might be to patch the CMake metadata here? Otherwise we'd have to patch the build system itself to ensure DLLs get installed into |
|
Actually, https://github.com/pytorch/pytorch/blob/v2.5.1/cmake/TorchConfig.cmake.in looks patchable |
|
I think (but did not test) that to install the c10 DLLs in bin it should be sufficient to remove the |
|
As a note, c10 is not the only issue; other targets' DLLs should also be set to bin, such as |
|
I'm working on a patch + test. |
Indeed, simply removing the destination does not work: |
Default DESTINATION argument is available for |
|
@traversaro, I've switched things to (only for TARGETS) -install(TARGETS torch_python DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+install(TARGETS torch_python
+ LIBRARY DESTINATION lib
+ RUNTIME DESTINATION bin)(and so on) but looking at the logs, it seems that the import libraries do not get installed anymore. You're right though, before I made the mistake of also changing some |
|
Either we delete DESTINATION from install(TARGETS, or add ARCHIVE DESTINATION for installing import libraries. |
gah, why would import libraries be handled by |
I get your pain, but just for reference the related docs is in https://cmake.org/cmake/help/latest/command/install.html#signatures . |
|
Even with the fixes from #318, there's still a problem with CMake, in that the CUDA-enabled builds (at least on linux), insist on finding actual CUDA drivers(?): This will obviously affect all feedstocks that want to build on top of a cuda-enabled pytorch. Someone else raised this in conda-forge/cuda-feedstock#59 already; the issue is that the Caffe2 CMake files still use the long-deprecated |
|
@danpetry, thanks for the offer to help. Here's my draft of replacing diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
index b51c7cc637b..6e107b5b02a 100644
--- a/caffe2/CMakeLists.txt
+++ b/caffe2/CMakeLists.txt
@@ -906,25 +906,25 @@ if(USE_ROCM)
"$<$<COMPILE_LANGUAGE:CXX>:ATen/core/ATen_pch.h>")
endif()
elseif(USE_CUDA)
- set(CUDA_LINK_LIBRARIES_KEYWORD PRIVATE)
+ set(CUDAToolkit_LINK_LIBRARIES_KEYWORD PRIVATE)
list(APPEND Caffe2_GPU_SRCS ${GENERATED_CXX_TORCH_CUDA})
- if(CUDA_SEPARABLE_COMPILATION)
+ if(CUDAToolkit_SEPARABLE_COMPILATION)
# Separate compilation fails when kernels using `thrust::sort_by_key`
# are linked with the rest of CUDA code. Workaround by linking them separately.
add_library(torch_cuda ${Caffe2_GPU_SRCS} ${Caffe2_GPU_CU_SRCS})
- set_property(TARGET torch_cuda PROPERTY CUDA_SEPARABLE_COMPILATION ON)
+ set_property(TARGET torch_cuda PROPERTY CUDAToolkit_SEPARABLE_COMPILATION ON)
add_library(torch_cuda_w_sort_by_key OBJECT
${Caffe2_GPU_SRCS_W_SORT_BY_KEY}
${Caffe2_GPU_CU_SRCS_W_SORT_BY_KEY})
- set_property(TARGET torch_cuda_w_sort_by_key PROPERTY CUDA_SEPARABLE_COMPILATION OFF)
+ set_property(TARGET torch_cuda_w_sort_by_key PROPERTY CUDAToolkit_SEPARABLE_COMPILATION OFF)
target_link_libraries(torch_cuda PRIVATE torch_cuda_w_sort_by_key)
else()
add_library(torch_cuda
${Caffe2_GPU_SRCS} ${Caffe2_GPU_SRCS_W_SORT_BY_KEY}
${Caffe2_GPU_CU_SRCS} ${Caffe2_GPU_CU_SRCS_W_SORT_BY_KEY})
endif()
- set(CUDA_LINK_LIBRARIES_KEYWORD)
+ set(CUDAToolkit_LINK_LIBRARIES_KEYWORD)
torch_compile_options(torch_cuda) # see cmake/public/utils.cmake
target_compile_definitions(torch_cuda PRIVATE USE_CUDA)
@@ -973,12 +973,12 @@ elseif(USE_CUDA)
torch_cuda
)
if($ENV{ATEN_STATIC_CUDA})
- if(CUDA_VERSION_MAJOR LESS_EQUAL 11)
+ if(CUDAToolkit_VERSION_MAJOR LESS_EQUAL 11)
target_link_libraries(torch_cuda_linalg PRIVATE
CUDA::cusolver_static
${CUDAToolkit_LIBRARY_DIR}/liblapack_static.a # needed for libcusolver_static
)
- elseif(CUDA_VERSION_MAJOR GREATER_EQUAL 12)
+ elseif(CUDAToolkit_VERSION_MAJOR GREATER_EQUAL 12)
target_link_libraries(torch_cuda_linalg PRIVATE
CUDA::cusolver_static
${CUDAToolkit_LIBRARY_DIR}/libcusolver_lapack_static.a # needed for libcusolver_static
diff --git a/cmake/Summary.cmake b/cmake/Summary.cmake
index d51c451589c..154f04a89dd 100644
--- a/cmake/Summary.cmake
+++ b/cmake/Summary.cmake
@@ -76,7 +76,7 @@ function(caffe2_print_configuration_summary)
message(STATUS " USE_CUSPARSELT : ${USE_CUSPARSELT}")
message(STATUS " USE_CUDSS : ${USE_CUDSS}")
message(STATUS " USE_CUFILE : ${USE_CUFILE}")
- message(STATUS " CUDA version : ${CUDA_VERSION}")
+ message(STATUS " CUDA version : ${CUDAToolkit_VERSION}")
message(STATUS " USE_FLASH_ATTENTION : ${USE_FLASH_ATTENTION}")
message(STATUS " USE_MEM_EFF_ATTENTION : ${USE_MEM_EFF_ATTENTION}")
if(${USE_CUDNN})
@@ -88,7 +88,7 @@ function(caffe2_print_configuration_summary)
if(${USE_CUFILE})
message(STATUS " cufile library : ${CUDA_cuFile_LIBRARY}")
endif()
- message(STATUS " CUDA root directory : ${CUDA_TOOLKIT_ROOT_DIR}")
+ message(STATUS " CUDA root directory : ${CUDAToolkit_ROOT}")
message(STATUS " CUDA library : ${CUDA_cuda_driver_LIBRARY}")
message(STATUS " cudart library : ${CUDA_cudart_LIBRARY}")
message(STATUS " cublas library : ${CUDA_cublas_LIBRARY}")
@@ -108,12 +108,12 @@ function(caffe2_print_configuration_summary)
message(STATUS " cuDSS library : ${__tmp}")
endif()
message(STATUS " nvrtc : ${CUDA_nvrtc_LIBRARY}")
- message(STATUS " CUDA include path : ${CUDA_INCLUDE_DIRS}")
- message(STATUS " NVCC executable : ${CUDA_NVCC_EXECUTABLE}")
+ message(STATUS " CUDA include path : ${CUDATookit_INCLUDE_DIRS}")
+ message(STATUS " NVCC executable : ${CUDATookit_NVCC_EXECUTABLE}")
message(STATUS " CUDA compiler : ${CMAKE_CUDA_COMPILER}")
message(STATUS " CUDA flags : ${CMAKE_CUDA_FLAGS}")
message(STATUS " CUDA host compiler : ${CMAKE_CUDA_HOST_COMPILER}")
- message(STATUS " CUDA --device-c : ${CUDA_SEPARABLE_COMPILATION}")
+ message(STATUS " CUDA --device-c : ${CUDATookit_SEPARABLE_COMPILATION}")
message(STATUS " USE_TENSORRT : ${USE_TENSORRT}")
if(${USE_TENSORRT})
message(STATUS " TensorRT runtime library: ${TENSORRT_LIBRARY}")
diff --git a/cmake/TorchConfig.cmake.in b/cmake/TorchConfig.cmake.in
index cba4d929855..da904fc6a18 100644
--- a/cmake/TorchConfig.cmake.in
+++ b/cmake/TorchConfig.cmake.in
@@ -125,7 +125,7 @@ if(@USE_CUDA@)
find_library(CAFFE2_NVRTC_LIBRARY caffe2_nvrtc PATHS "${TORCH_INSTALL_PREFIX}/lib")
list(APPEND TORCH_CUDA_LIBRARIES ${CAFFE2_NVRTC_LIBRARY})
else()
- set(TORCH_CUDA_LIBRARIES ${CUDA_NVRTC_LIB})
+ set(TORCH_CUDA_LIBRARIES CUDA::nvrtc)
endif()
if(TARGET torch::nvtoolsext)
list(APPEND TORCH_CUDA_LIBRARIES torch::nvtoolsext)
diff --git a/cmake/public/cuda.cmake b/cmake/public/cuda.cmake
index 152fbdbe6dd..13bae9b6227 100644
--- a/cmake/public/cuda.cmake
+++ b/cmake/public/cuda.cmake
@@ -26,8 +26,8 @@ if(NOT MSVC)
endif()
# Find CUDA.
-find_package(CUDA)
-if(NOT CUDA_FOUND)
+find_package(CUDAToolkit)
+if(NOT CUDAToolkit_FOUND)
message(WARNING
"Caffe2: CUDA cannot be found. Depending on whether you are building "
"Caffe2 or a Caffe2 dependent library, the next warning / error will "
@@ -36,8 +36,6 @@ if(NOT CUDA_FOUND)
return()
endif()
-# Enable CUDA language support
-set(CUDAToolkit_ROOT "${CUDA_TOOLKIT_ROOT_DIR}")
# Pass clang as host compiler, which according to the docs
# Must be done before CUDA language is enabled, see
# https://cmake.org/cmake/help/v3.15/variable/CMAKE_CUDA_HOST_COMPILER.html
@@ -56,24 +54,18 @@ if(CMAKE_VERSION VERSION_GREATER_EQUAL 3.12.0)
cmake_policy(SET CMP0074 NEW)
endif()
-find_package(CUDAToolkit REQUIRED)
+find_package(CUDAToolkit REQUIRED COMPONENTS cudart nvrtc REQUIRED)
cmake_policy(POP)
-if(NOT CMAKE_CUDA_COMPILER_VERSION VERSION_EQUAL CUDAToolkit_VERSION)
- message(FATAL_ERROR "Found two conflicting CUDA versions:\n"
- "V${CMAKE_CUDA_COMPILER_VERSION} in '${CUDA_INCLUDE_DIRS}' and\n"
- "V${CUDAToolkit_VERSION} in '${CUDAToolkit_INCLUDE_DIRS}'")
-endif()
-
-message(STATUS "Caffe2: CUDA detected: " ${CUDA_VERSION})
-message(STATUS "Caffe2: CUDA nvcc is: " ${CUDA_NVCC_EXECUTABLE})
-message(STATUS "Caffe2: CUDA toolkit directory: " ${CUDA_TOOLKIT_ROOT_DIR})
-if(CUDA_VERSION VERSION_LESS 11.0)
+message(STATUS "Caffe2: CUDA detected: " ${CUDAToolkit_VERSION})
+message(STATUS "Caffe2: CUDA nvcc is: " ${CUDAToolkit_NVCC_EXECUTABLE})
+message(STATUS "Caffe2: CUDA toolkit directory: " ${CUDAToolkit_ROOT})
+if(CUDAToolkit_VERSION VERSION_LESS 11.0)
message(FATAL_ERROR "PyTorch requires CUDA 11.0 or above.")
endif()
-if(CUDA_FOUND)
+if(CUDAToolkit_FOUND)
# Sometimes, we may mismatch nvcc with the CUDA headers we are
# compiling with, e.g., if a ccache nvcc is fed to us by CUDA_NVCC_EXECUTABLE
# but the PATH is not consistent with CUDA_HOME. It's better safe
@@ -97,8 +89,8 @@ if(CUDA_FOUND)
)
if(NOT CMAKE_CROSSCOMPILING)
try_run(run_result compile_result ${PROJECT_RANDOM_BINARY_DIR} ${file}
- CMAKE_FLAGS "-DINCLUDE_DIRECTORIES=${CUDA_INCLUDE_DIRS}"
- LINK_LIBRARIES ${CUDA_LIBRARIES}
+ CMAKE_FLAGS "-DINCLUDE_DIRECTORIES=${CUDAToolkit_INCLUDE_DIRS}"
+ LINK_LIBRARIES ${CUDAToolkit_LIBRARIES}
RUN_OUTPUT_VARIABLE cuda_version_from_header
COMPILE_OUTPUT_VARIABLE output_var
)
@@ -106,20 +98,20 @@ if(CUDA_FOUND)
message(FATAL_ERROR "Caffe2: Couldn't determine version from header: " ${output_var})
endif()
message(STATUS "Caffe2: Header version is: " ${cuda_version_from_header})
- if(NOT cuda_version_from_header STREQUAL ${CUDA_VERSION_STRING})
+ if(NOT cuda_version_from_header STREQUAL ${CUDAToolkit_VERSION_STRING})
# Force CUDA to be processed for again next time
# TODO: I'm not sure if this counts as an implementation detail of
# FindCUDA
- set(${cuda_version_from_findcuda} ${CUDA_VERSION_STRING})
+ set(${cuda_version_from_findcuda} ${CUDAToolkit_VERSION_STRING})
unset(CUDA_TOOLKIT_ROOT_DIR_INTERNAL CACHE)
# Not strictly necessary, but for good luck.
- unset(CUDA_VERSION CACHE)
+ unset(CUDAToolkit_VERSION CACHE)
# Error out
message(FATAL_ERROR "FindCUDA says CUDA version is ${cuda_version_from_findcuda} (usually determined by nvcc), "
"but the CUDA headers say the version is ${cuda_version_from_header}. This often occurs "
"when you set both CUDA_HOME and CUDA_NVCC_EXECUTABLE to "
"non-standard locations, without also setting PATH to point to the correct nvcc. "
- "Perhaps, try re-running this command again with PATH=${CUDA_TOOLKIT_ROOT_DIR}/bin:$PATH. "
+ "Perhaps, try re-running this command again with PATH=${CUDAToolkit_ROOT}/bin:$PATH. "
"See above log messages for more diagnostics, and see https://github.com/pytorch/pytorch/issues/8092 for more details.")
endif()
endif()
@@ -128,8 +120,8 @@ endif()
# ---[ CUDA libraries wrapper
# find lbnvrtc.so
-set(CUDA_NVRTC_LIB "${CUDA_nvrtc_LIBRARY}" CACHE FILEPATH "")
-if(CUDA_NVRTC_LIB AND NOT CUDA_NVRTC_SHORTHASH)
+get_target_property(CUDA_NVRTC_LIB cuda::nvrtc INTERFACE_LINK_LIBRARIES)
+if(NOT CUDA_NVRTC_SHORTHASH)
find_package(Python COMPONENTS Interpreter)
execute_process(
COMMAND Python::Interpreter -c |
|
looks like you've removed |
That should be set by |
Solution to issue cannot be found in the documentation.
Issue
Description:
The
libtorchpackage for Windows in CMake is currently not working as expected. When usingfind_package(Torch)in aCMakeLists.txtfile, it returns an error due to missing .dll files:Reproduction Steps:
Create a new CMake project with the following CMakeLists.txt:
Run cmake on the project and observe the error message.
Note: This issue only occurs on Windows, while macOS and Linux work fine.
The
.dllfiles are located in<prefix>/Library/bin/*, but the configuration is set to find them in<prefix>/Library/lib/*. I'm unsure where to look into to resolve this issue. Any help would be appreciated!Installed packages
Environment info
System ------------ Pixi version: 0.40.3 Platform: win-64 Virtual packages: __win=0=0 : __cuda=12.5=0 : __archspec=1=x86_64_v3 Cache dir: C:\Users\jeongseok\AppData\Local\rattler\cache Auth storage: C:\Users\jeongseok\.rattler\credentials.json Config locations: No config files found Global ------------ Bin dir: <root>\.pixi\bin Environment dir: <root>\.pixi\envs Manifest dir: <root>\.pixi\manifests\pixi-global.toml Project ------------ Name: test_libtorch Version: 0.1.0 Manifest file: <root>\test_libtorch\pixi.toml Last updated: 27-01-2025 12:58:50 Environments ------------ Environment: default Features: default Channels: conda-forge Dependency count: 1 Dependencies: libtorch Target platforms: win-64 Tasks: configThe text was updated successfully, but these errors were encountered: