[ERROR] in the last step of pip install .

[wplf]

Bug

/TransformerEngine/transformer_engine/common/util/cuda_driver.cpp:98:3: error: ‘cudaDriverEntryPointQueryResult’ was not declared in this scope

This bug may be caused by commit "f9dd37f7bd4c08c79bd6b7bece25dfe286225458"

Sadly, I have to develop on v1.7.

Environment

• torch 2.4
• CUDA 118
• NVCC 11.8
• GCC 9.4.0
• cmake 3.16.3

Description

The building process stopped at the last step [32/33]

[32/33] Building CUDA object CMakeFiles/transformer_engine.dir/layer_norm/ln_fwd_cuda_kernel.cu.o ninja: build stopped: subcommand failed.

/vepfs/home/lijinliang/miniconda3/envs/megatron/lib/python3.10/site-packages/cmake/data/bin/cmake --build /vepfs/home/lijinliang/projects/TransformerEngine/build/cmake --parallel 32
[1/2] Building CXX object CMakeFiles/transformer_engine.dir/util/cuda_driver.cpp.o
FAILED: CMakeFiles/transformer_engine.dir/util/cuda_driver.cpp.o
/usr/bin/c++ -DNV_CUDNN_FRONTEND_USE_DYNAMIC_LOADING -Dtransformer_engine_EXPORTS -I/vepfs/home/lijinliang/projects/TransformerEngine/transformer_engine/common/.. -I/vepfs/home/lijinliang/projects/TransformerEngine/transformer_engine/common/include -I/vepfs/home/lijinliang/projects/TransformerEngine/transformer_engine/common/../../3rdparty/cudnn-frontend/include -I/vepfs/home/lijinliang/projects/TransformerEngine/build/cmake/string_headers -isystem /usr/local/cuda/targets/x86_64-linux/include -O3 -DNDEBUG -std=gnu++17 -fPIC -MD -MT CMakeFiles/transformer_engine.dir/util/cuda_driver.cpp.o -MF CMakeFiles/transformer_engine.dir/util/cuda_driver.cpp.o.d -o CMakeFiles/transformer_engine.dir/util/cuda_driver.cpp.o -c /vepfs/home/lijinliang/projects/TransformerEngine/transformer_engine/common/util/cuda_driver.cpp
/vepfs/home/lijinliang/projects/TransformerEngine/transformer_engine/common/util/cuda_driver.cpp: In function ‘void* transformer_engine::cuda_driver::get_symbol(const char*)’:
/vepfs/home/lijinliang/projects/TransformerEngine/transformer_engine/common/util/cuda_driver.cpp:98:3: error: ‘cudaDriverEntryPointQueryResult’ was not declared in this scope
98 | cudaDriverEntryPointQueryResult driver_result;
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In file included from /vepfs/home/lijinliang/projects/TransformerEngine/transformer_engine/common/util/../common.h:25,
from /vepfs/home/lijinliang/projects/TransformerEngine/transformer_engine/common/util/cuda_driver.cpp:11:
/vepfs/home/lijinliang/projects/TransformerEngine/transformer_engine/common/util/cuda_driver.cpp:99:85: error: ‘driver_result’ was not declared in this scope
99 | NVTE_CHECK_CUDA(cudaGetDriverEntryPoint(symbol, &entry_point, cudaEnableDefault, &driver_result));
| ^~~~~~~~~~~~~
/vepfs/home/lijinliang/projects/TransformerEngine/transformer_engine/common/util/.././util/logging.h:36:49: note: in definition of macro ‘NVTE_CHECK_CUDA’
36 | const cudaError_t status_NVTE_CHECK_CUDA = (expr);
| ^~~~
/vepfs/home/lijinliang/projects/TransformerEngine/transformer_engine/common/util/cuda_driver.cpp:100:14: error: ‘driver_result’ was not declared in this scope
100 | NVTE_CHECK(driver_result == cudaDriverEntryPointSuccess,
| ^~~~~~~~~~~~~
/vepfs/home/lijinliang/projects/TransformerEngine/transformer_engine/common/util/.././util/logging.h:28:11: note: in definition of macro ‘NVTE_CHECK’
28 | if (!(expr)) {
| ^~~~
/vepfs/home/lijinliang/projects/TransformerEngine/transformer_engine/common/util/cuda_driver.cpp:100:31: error: ‘cudaDriverEntryPointSuccess’ was not declared in this scope; did you mean ‘cudaGetDriverEntryPointFlags’?
100 | NVTE_CHECK(driver_result == cudaDriverEntryPointSuccess,
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~
/vepfs/home/lijinliang/projects/TransformerEngine/transformer_engine/common/util/.././util/logging.h:28:11: note: in definition of macro ‘NVTE_CHECK’
28 | if (!(expr)) {
| ^~~~
ninja: build stopped: subcommand failed.

[wplf]

I am wondering if the cuda lib causes this problem, but NVCC and cuda lib is in my path and LD_LIBRARY_PATH.

[timmoon10]

Can you try upgrading to CUDA 12.0 or newer?

This error shows up because #970 uses cudaGetDriverEntryPoint to access the CUDA driver. However, the function signature is slightly different in CUDA 11.8 and CUDA 12.0. We should either bump the minimum CUDA version or add some version-dependent logic at:

TransformerEngine/transformer_engine/common/util/cuda_driver.cpp

Lines 96 to 103 in 6717554

	void *get_symbol(const char *symbol) {
	void *entry_point;
	cudaDriverEntryPointQueryResult driver_result;
	NVTE_CHECK_CUDA(cudaGetDriverEntryPoint(symbol, &entry_point, cudaEnableDefault, &driver_result));
	NVTE_CHECK(driver_result == cudaDriverEntryPointSuccess,
	"Could not find CUDA driver entry point for ", symbol);
	return entry_point;
	}

[wplf]

Thanks you for your kindness.
I will upgrade cuda!

[jinghere11]

For I can't upgrade CUDA with a sharing env, can you tell me which version of the Transformer Engine could work?

[wplf]

Of course, my version is 1.7.0+4e7caa1, you can check out 4e7caa1, and build from source.